<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: AI Release</title>
    <description>The latest articles on DEV Community by AI Release (@qunirice_verase_1ab0caeba).</description>
    <link>https://dev.to/qunirice_verase_1ab0caeba</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4206699%2F3c8087a2-fa91-4108-8b20-80339144b365.png</url>
      <title>DEV Community: AI Release</title>
      <link>https://dev.to/qunirice_verase_1ab0caeba</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/qunirice_verase_1ab0caeba"/>
    <language>en</language>
    <item>
      <title>Hybrid LLM routing on iPhone: local models plus cloud APIs with Shortcuts</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Mon, 16 Nov 2026 10:13:31 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/hybrid-llm-routing-on-iphone-local-models-plus-cloud-apis-with-shortcuts-26gj</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/hybrid-llm-routing-on-iphone-local-models-plus-cloud-apis-with-shortcuts-26gj</guid>
      <description>&lt;p&gt;Most of the prompts I type on my phone are trivial — a quick rewrite, a two-line summary, a "what does this mean" question. Sending all of those to a paid cloud endpoint felt wasteful, and some of them contained things I would rather not upload at all. So I built a small router in Apple Shortcuts that keeps the light and sensitive requests on a local model and only sends the heavy lifting to a cloud API.&lt;/p&gt;

&lt;p&gt;No Xcode, no full iOS app. Just Shortcuts, an HTTP request, and one length check.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this setup actually does
&lt;/h2&gt;

&lt;p&gt;The idea is simple: one shortcut, two branches.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Short or sensitive prompts go to a local model running on the iPhone itself.&lt;/li&gt;
&lt;li&gt;Long, complex, or "I need the good model" prompts go to a cloud provider.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Everything happens inside Shortcuts, Apple's built-in automation app, which chains actions into a single runnable script. You can trigger it from the app, from Siri, or from the Share Sheet.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prerequisites and a pre-flight checklist
&lt;/h2&gt;

&lt;p&gt;Before touching Shortcuts, make sure the pieces talk to each other:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;iPhone with iOS 17 or later.&lt;/strong&gt; Shortcuts ships with the system. Verify under Settings → General → About → Software Version.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A local AI app&lt;/strong&gt; that runs a model on the device or exposes a local HTTP endpoint — meaning a URL on your phone that accepts requests, such as &lt;code&gt;http://127.0.0.1:8080/v1/chat/completions&lt;/code&gt;. The port depends on your app.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A cloud provider account and API key.&lt;/strong&gt; The key is the secret string (often starting with &lt;code&gt;sk-&lt;/code&gt;) that identifies your account.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Internet access&lt;/strong&gt; for the cloud branch. The local branch can work offline once the model is downloaded.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A routing rule&lt;/strong&gt; you can state in one sentence.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;My checklist before I started:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;[ ] Shortcuts opens and I can create a new shortcut.&lt;/li&gt;
&lt;li&gt;[ ] My local app answers in Safari at the local URL.&lt;/li&gt;
&lt;li&gt;[ ] My cloud API key works when tested from a desktop (curl or the provider console).&lt;/li&gt;
&lt;li&gt;[ ] I know the exact JSON field names each provider expects.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That last item matters more than it sounds. Two "OpenAI-compatible" endpoints can still disagree on field names, and you will only find out at runtime.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 1: Choose the routing rule first
&lt;/h2&gt;

&lt;p&gt;Write the rule down before you build anything. Mine looked like this:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Prompts shorter than 200 characters → local.&lt;/li&gt;
&lt;li&gt;Prompts containing "private" or "confidential" → local, no exceptions.&lt;/li&gt;
&lt;li&gt;Everything else → cloud.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The 200-character threshold is a heuristic, not science. Short prompts are usually cheap and fast enough for a small on-device model. If yours handles more, raise the number.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: Create the shortcut and the length check
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Open Shortcuts, tap &lt;strong&gt;+&lt;/strong&gt;, and name it &lt;code&gt;Hybrid AI Router&lt;/code&gt;. You get an empty canvas.&lt;/li&gt;
&lt;li&gt;Add a &lt;strong&gt;Text&lt;/strong&gt; action and put a sample prompt in it, for example: &lt;code&gt;Summarize this note in three bullets: The meeting moved to Friday.&lt;/code&gt; Later you can swap this for Shortcut Input when you run it from Siri or the Share Sheet.&lt;/li&gt;
&lt;li&gt;Add &lt;strong&gt;Get Length of Text&lt;/strong&gt; and point it at the Text action.&lt;/li&gt;
&lt;li&gt;Add an &lt;strong&gt;If&lt;/strong&gt; action: &lt;code&gt;Length&lt;/code&gt; → &lt;code&gt;is less than&lt;/code&gt; → &lt;code&gt;200&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;You now have two branches. A prompt like &lt;code&gt;What is 2+2?&lt;/code&gt; is 8 characters, so it falls into the local branch. A request for a 500-word essay lands in the cloud branch. That is the whole routing logic — everything else is plumbing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: Build the local branch
&lt;/h2&gt;

&lt;p&gt;Inside the If branch:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Add &lt;strong&gt;Get Contents of URL&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Set the URL to your local endpoint, e.g. &lt;code&gt;http://127.0.0.1:8080/v1/chat/completions&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Set Method to &lt;code&gt;POST&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Add the header &lt;code&gt;Content-Type: application/json&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Then add a &lt;strong&gt;Request Body&lt;/strong&gt; in JSON. This template works for most local runtimes, since they tend to speak the OpenAI-compatible chat format:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"model"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"your-local-model"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"messages"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
    &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"role"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"user"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"content"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"What is 2+2?"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"temperature"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mf"&gt;0.7&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Replace &lt;code&gt;your-local-model&lt;/code&gt; with the model name your app reports. If your app ignores the model field, delete that line entirely. A correct run ends with your local app returning a JSON response that the shortcut can read.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 4: Build the cloud branch
&lt;/h2&gt;

&lt;p&gt;In the &lt;strong&gt;Otherwise&lt;/strong&gt; branch, add a second &lt;strong&gt;Get Contents of URL&lt;/strong&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;URL: your provider's chat completion endpoint, e.g. &lt;code&gt;https://api.example.com/v1/chat/completions&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Method: &lt;code&gt;POST&lt;/code&gt;, header &lt;code&gt;Content-Type: application/json&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Body: the same JSON shape, but with your cloud model name in the &lt;code&gt;model&lt;/code&gt; field and the actual prompt in &lt;code&gt;content&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Authentication: add the header exactly as your provider's docs describe it. The key usually goes into a request header, but the header name differs between vendors — check the docs rather than guessing.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;One practical note: keeping the API key inside a shortcut is convenient but not a vault. Anyone with access to your unlocked phone can open the shortcut and read it. Treat a shortcut key as a low-limit throwaway credential, and rotate it if the phone is shared.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 5: Test both branches on purpose
&lt;/h2&gt;

&lt;p&gt;Do not trust a single happy-path run. I tested by editing the Text action:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A 5-character prompt to confirm the local branch fires.&lt;/li&gt;
&lt;li&gt;A 400-character prompt to confirm the cloud branch fires.&lt;/li&gt;
&lt;li&gt;A short prompt containing the word "confidential" to confirm the keyword rule overrides the length rule.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If the local branch fails, open the local URL in Safari first — if Safari cannot reach it, Shortcuts will not either. If the cloud branch fails, re-test the key from a desktop curl call to separate an auth problem from a shortcut problem.&lt;/p&gt;

&lt;p&gt;Two things I would change next time: switching the Text action to Shortcut Input so the router works from Siri and the Share Sheet, and adding a fallback so a failed cloud call retries locally instead of just erroring out.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;You do not need a custom app to get hybrid AI on an iPhone. A length check and two HTTP actions in Shortcuts are enough to keep cheap, private prompts on-device and spend cloud tokens only where they buy you something. Start with a crude rule like 200 characters, run both branches deliberately, and tune the threshold based on what your local model actually handles well. If you are still choosing which local model to run, AI Release also has beginner guides for running models locally on a PC and on Android — worth a read before you hardcode a gateway.&lt;/p&gt;

&lt;p&gt;This article was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;More practical AI guides and breakdowns: &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>apple</category>
      <category>tutorial</category>
      <category>llm</category>
    </item>
    <item>
      <title>Running a Local LLM Chat on Android: Apps, Termux, and What to Expect</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Mon, 16 Nov 2026 10:07:29 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/running-a-local-llm-chat-on-android-apps-termux-and-what-to-expect-4a10</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/running-a-local-llm-chat-on-android-apps-termux-and-what-to-expect-4a10</guid>
      <description>&lt;p&gt;I wanted to chat with a language model on my phone without shipping every message to someone's server. It turned out I did not need root, a flagship device, or a cloud account - just enough free storage and a bit of patience. Here is what I learned while trying the different ways to get an LLM running locally on Android.&lt;/p&gt;

&lt;h2&gt;
  
  
  What a local AI companion actually is
&lt;/h2&gt;

&lt;p&gt;A local companion is simply an app that runs the language model on the phone itself instead of calling a remote API. The model is the neural network that writes the replies. Because it lives on your device, the conversation stays there and keeps working in airplane mode.&lt;/p&gt;

&lt;p&gt;Most of these apps support model families such as Gemma, Llama, Phi or Mistral. Specific names come and go, but the workflow does not change: install the app, download a model file, start chatting. That is the whole loop.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check your phone before you download anything
&lt;/h2&gt;

&lt;p&gt;I skipped this step the first time and wasted an evening. A few numbers matter:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Android version:&lt;/strong&gt; 8.0 or newer covers most tools.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;RAM:&lt;/strong&gt; 4 GB is the floor; 6-8 GB is noticeably smoother.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Storage:&lt;/strong&gt; model files run from about 1 GB up to 10 GB or more. Free space first.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Battery:&lt;/strong&gt; local inference is heavier than a normal chat app, so expect drain.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;App source:&lt;/strong&gt; install from Google Play or the project's official page. Random APK files are not worth the risk.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A practical example: if Settings already warns you about low storage, a 2 GB model will not install on a phone with 1.5 GB free. Clear videos and unused apps before you start the download, not after it fails.&lt;/p&gt;

&lt;h2&gt;
  
  
  Route 1: ready-made companion apps
&lt;/h2&gt;

&lt;p&gt;These give you a chat window, a "download model" button and a small settings page. For most people this is the right starting point - fast setup, decent interface, less control over the technical details. Some are free, some ask for a one-time payment.&lt;/p&gt;

&lt;p&gt;In practice you install the app, open Settings, tap the model download option and pick a small version. The progress bar usually starts right away, and the chat becomes usable once it finishes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Route 2: Termux plus llama.cpp
&lt;/h2&gt;

&lt;p&gt;Termux is a terminal emulator that gives you a Linux-like environment on Android without root. &lt;code&gt;llama.cpp&lt;/code&gt; runs language models on ordinary CPUs. Together they are the tinkerer's path, and you do need to be comfortable typing commands.&lt;/p&gt;

&lt;p&gt;A minimal session looks like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pkg install llama-cpp
llama-cli -m /sdcard/Download/model.gguf -p &lt;span class="s2"&gt;"Hello, how are you?"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The &lt;code&gt;-p&lt;/code&gt; flag sets the opening prompt, and the terminal prints the model's answer after a few seconds. It works because &lt;code&gt;llama.cpp&lt;/code&gt; is compiled for low-resource hardware and Termux provides the missing environment. The upside is full control. The downside is that everything - model choice, parameters, file paths - is on you.&lt;/p&gt;

&lt;h2&gt;
  
  
  Route 3: cloud with a local fallback
&lt;/h2&gt;

&lt;p&gt;Hybrid apps behave like normal cloud chat when you have a connection and switch to an on-device model when you do not. That suits people who want fast answers on Wi-Fi and private conversations on the move.&lt;/p&gt;

&lt;p&gt;Setup is usually a single toggle in the settings, something like "Prefer offline mode" or "Local model". Just be aware that the moment cloud mode is active, your messages may reach the provider's servers.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I choose between them
&lt;/h2&gt;

&lt;p&gt;I compare four things:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Model support.&lt;/strong&gt; Can you pick the model size and format yourself? GGUF is the common format for local models.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;True offline mode.&lt;/strong&gt; Does the chat work in airplane mode after the first download?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Memory.&lt;/strong&gt; Does the app keep earlier messages in context? Look for "context length" or "memory" in the description.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interface.&lt;/strong&gt; Start with the free tier. You will be living in that chat window.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If a developer says "supports GGUF", you can load models in that format. If an app accepts one fixed model only, ask yourself whether that model fits your phone before committing.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Approach&lt;/th&gt;
&lt;th&gt;Best for&lt;/th&gt;
&lt;th&gt;Typical device&lt;/th&gt;
&lt;th&gt;Price&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Ready-made companion app&lt;/td&gt;
&lt;td&gt;Beginners, quick private chat&lt;/td&gt;
&lt;td&gt;6-8 GB RAM, 5-10 GB storage&lt;/td&gt;
&lt;td&gt;Free or one-time payment&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Termux + llama.cpp&lt;/td&gt;
&lt;td&gt;Enthusiasts who want control&lt;/td&gt;
&lt;td&gt;8 GB RAM, 10+ GB storage&lt;/td&gt;
&lt;td&gt;Free (open source)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cloud + local fallback&lt;/td&gt;
&lt;td&gt;People who need both modes&lt;/td&gt;
&lt;td&gt;4-6 GB RAM, 2-3 GB storage&lt;/td&gt;
&lt;td&gt;Subscription or freemium&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  A 10-minute first test
&lt;/h2&gt;

&lt;p&gt;This works with most ready-made apps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Install the app from Google Play. You should land on a welcome screen offering a model download.&lt;/li&gt;
&lt;li&gt;Open the model download section in Settings. You should see a list of model sizes.&lt;/li&gt;
&lt;li&gt;Pick the smallest one, usually under 1 GB. The download starts, and the chat unlocks when it completes.&lt;/li&gt;
&lt;li&gt;Turn on airplane mode and send a short greeting. The model should answer without any mobile connection.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Starting small keeps the risk of crashes low. Once the tiny model replies, try a bigger one and compare the quality - you will immediately feel the difference in coherence and speed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Mistakes I hit, and the fixes
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;The app crashes right after loading the model.&lt;/strong&gt; The screen freezes or the app closes on the first message. Fix: download a smaller variant. Many model pages offer 1 GB, 3 GB and 5 GB versions of the same family.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The download stalls around 90%.&lt;/strong&gt; You get an "Error downloading model" message. Fix: delete old videos or unused apps and restart. The model needs its full size available at once, not in fragments.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;"Unsupported model file".&lt;/strong&gt; The file is there but the app refuses to open it. Fix: grab a GGUF model and choose a common quantization such as &lt;code&gt;Q4_K_M&lt;/code&gt;. That pairing works with most Android tools.&lt;/p&gt;

&lt;h2&gt;
  
  
  Quick answers
&lt;/h2&gt;

&lt;p&gt;&lt;em&gt;Completely offline?&lt;/em&gt; Yes - internet is only needed to download the app and the model file. After that, airplane mode is fine.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Only 4 GB of RAM?&lt;/em&gt; You can still do it, but stick to small models, expect slower replies and close other apps while chatting.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;How much space?&lt;/em&gt; Small models sit under 1 GB, medium ones take 2-5 GB, large ones go past 10 GB. Check the model page before hitting download.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Is it private?&lt;/em&gt; With a fully offline app your messages stay on the device. If cloud fallback is enabled, they may go to the provider's server.&lt;/p&gt;

&lt;p&gt;This article was originally published on AI Release - &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt;. More notes and updates land on the channel &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;A local AI companion on Android is not exotic anymore - it is a download and a settings tap. Match the model size to your RAM and free storage, test with the smallest model first in airplane mode, and only then move up. Ready-made apps get you there in minutes; Termux and &lt;code&gt;llama.cpp&lt;/code&gt; give you control if you want it.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>android</category>
      <category>llm</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>Local RAG on a mini PC: EmbeddingGemma 2, PDF Desk, and when to add a vector store</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Sun, 15 Nov 2026 10:12:29 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/local-rag-on-a-mini-pc-embeddinggemma-2-pdf-desk-and-when-to-add-a-vector-store-3ia7</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/local-rag-on-a-mini-pc-embeddinggemma-2-pdf-desk-and-when-to-add-a-vector-store-3ia7</guid>
      <description>&lt;p&gt;Local RAG doesn't require a GPU tower humming under your desk. On a mini PC you can assemble a working retrieval pipeline from two published pieces: EmbeddingGemma 2, an open 740M-parameter embedding model, and PDF Desk, a small Python/Streamlit app that answers questions about a PDF and points at the page it used. Vector stores like Qdrant or Chroma are the optional next step — not a prerequisite for getting something running today.&lt;/p&gt;

&lt;p&gt;Here's the path, step by step, with the expected result at each stage and the places where it usually goes sideways.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you need before you start
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;A mini PC with enough RAM for a 740M-parameter model. Check this first — it's the one thing &lt;code&gt;pip&lt;/code&gt; can't fix.&lt;/li&gt;
&lt;li&gt;A Python environment for PDF Desk.&lt;/li&gt;
&lt;li&gt;macOS or Linux. On Windows the command variants live in the repo README, so the commands below won't line up.&lt;/li&gt;
&lt;li&gt;Hugging Face access, to pull the EmbeddingGemma 2 weights.&lt;/li&gt;
&lt;li&gt;The PDF Desk repository from GitHub.&lt;/li&gt;
&lt;li&gt;One test PDF. PDF Desk handles a single active document at a time.&lt;/li&gt;
&lt;li&gt;Optional: Qdrant or Chroma if you plan to index many documents or non-PDF content.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A checklist I'd actually tick off:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;[ ] EmbeddingGemma 2 weights downloaded from Hugging Face&lt;/li&gt;
&lt;li&gt;[ ] PDF Desk source cloned from GitHub&lt;/li&gt;
&lt;li&gt;[ ] Python and dependencies installed&lt;/li&gt;
&lt;li&gt;[ ] One test PDF ready&lt;/li&gt;
&lt;li&gt;[ ] Free RAM and disk space confirmed on the mini PC&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Step 1: Pull EmbeddingGemma 2 from Hugging Face
&lt;/h2&gt;

&lt;p&gt;EmbeddingGemma 2 is Google DeepMind's open multimodal embedding model at 740 million parameters, and it maps text, code, images, video and audio into one vector space. That's the part that matters for RAG: your query and your document chunks land in the same space, so comparing them is a straightforward similarity lookup.&lt;/p&gt;

&lt;p&gt;Find the model page on Hugging Face and download the weights into a folder you can point at later. Expected result: model files sitting locally on the mini PC, no API key involved. If the download stalls, check connectivity and confirm the repository is reachable — big files also need free disk space and headroom in RAM.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: Launch PDF Desk and ask your first question
&lt;/h2&gt;

&lt;p&gt;PDF Desk is a Python application with a Streamlit front end. Clone it, install dependencies, run it, and the interface opens in your browser.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone &lt;span class="o"&gt;&amp;lt;&lt;/span&gt;pdf-desk-repo-url&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;      &lt;span class="c"&gt;# exact URL is on the GitHub page&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; pdf-desk
python -m venv .venv
&lt;span class="nb"&gt;source&lt;/span&gt; .venv/bin/activate          &lt;span class="c"&gt;# Windows users: follow the README&lt;/span&gt;
pip install -r requirements.txt
streamlit run &lt;span class="o"&gt;&amp;lt;&lt;/span&gt;entry-file&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;         &lt;span class="c"&gt;# file name as listed in the README&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then upload one PDF, type a question, and read the answer. The detail that makes this useful is the page reference: the response comes back tied to a specific page number, so you can verify the claim instead of trusting the prose. When you need a different document, replace the active file and re-upload before asking the next question.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: Decide whether Qdrant or Chroma earns its place
&lt;/h2&gt;

&lt;p&gt;Neither source post walks through a vector store installation, so I won't invent one — for setup details, the official Qdrant and Chroma documentation is the place to go. What the posts do make clear is the division of labour:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Your situation&lt;/th&gt;
&lt;th&gt;What's enough&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;One PDF, page-level answers&lt;/td&gt;
&lt;td&gt;PDF Desk indexes it itself&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Many documents, or images/audio/code&lt;/td&gt;
&lt;td&gt;EmbeddingGemma 2 vectors plus Qdrant or Chroma&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The pipeline shape is identical either way: EmbeddingGemma 2 turns content into vectors, the store persists them, and queries are embedded into the same space so semantic search works. PDF Desk is the shortest path to a demo; a vector database is what you reach for once "one active PDF" becomes the blocker.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it breaks, and the quick fix
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Only one active PDF at a time.&lt;/strong&gt; That's a documented limitation. Swap the file and re-upload; don't expect a folder-wide index.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No Windows commands in the walkthrough.&lt;/strong&gt; Open the GitHub README and use the Windows section instead of copy-pasting macOS/Linux lines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hugging Face download fails.&lt;/strong&gt; Verify the network, confirm you can reach the repository, and free up disk and RAM for a model this size.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;You want everything indexed.&lt;/strong&gt; That's the vector store's job, not PDF Desk's.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Quick answers
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is EmbeddingGemma 2?&lt;/strong&gt; An open multimodal embedding model from Google DeepMind with 740M parameters, combining text, code, images, video and audio in one vector space.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can a mini PC run local RAG?&lt;/strong&gt; Yes — EmbeddingGemma 2 is open on Hugging Face and PDF Desk runs locally on macOS/Linux, and both are light enough for a mini PC with sufficient RAM.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Qdrant or Chroma — which one?&lt;/strong&gt; The source posts don't compare them. For plain PDF search, PDF Desk indexes on its own; for a bigger RAG build, pick based on your preference and follow the official setup guides.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do page references work?&lt;/strong&gt; PDF Desk links each answer to the relevant page of the active document, which is the core feature described in the original post.&lt;/p&gt;

&lt;p&gt;If you want the original write-ups and more local-AI material, this article was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt; — and the channel posts are collected at &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;Start small: one PDF, PDF Desk, page-cited answers — a fully local RAG loop you can validate in an afternoon. Bring in EmbeddingGemma 2 when you want a custom embedding space, and only then reach for Qdrant or Chroma, once the one-document limit is the thing actually slowing you down.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>tutorial</category>
      <category>python</category>
    </item>
    <item>
      <title>LoRA fine-tuning on one GPU: from JSONL to a working adapter</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Sun, 15 Nov 2026 10:06:28 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/lora-fine-tuning-on-one-gpu-from-jsonl-to-a-working-adapter-2706</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/lora-fine-tuning-on-one-gpu-from-jsonl-to-a-working-adapter-2706</guid>
      <description>&lt;p&gt;I wanted a model that answered in my own format — not a general chatbot, just something that reliably turned messy input into the labels I cared about. Retraining a full model was out of reach, but LoRA on a single consumer card turned out to be very doable. Here is the process I followed, including the parts that broke.&lt;/p&gt;

&lt;p&gt;This is written for someone with one GPU (8–24 GB of VRAM), a Linux or Windows box, and Python installed. You don't need a cluster or a big budget. You do need to be comfortable with a terminal.&lt;/p&gt;

&lt;h2&gt;
  
  
  What LoRA actually changes
&lt;/h2&gt;

&lt;p&gt;LoRA keeps the base model frozen and bolts small adapter matrices onto selected layers. During training, only those adapters update, which is why the memory footprint stays small: you're not storing gradients and optimizer states for billions of weights, only for a thin slice of them.&lt;/p&gt;

&lt;p&gt;The result is a checkpoint measured in tens of megabytes instead of gigabytes. You keep the base model around and load the adapter on top of it at inference time.&lt;/p&gt;

&lt;p&gt;The models that fit this workflow comfortably are small language models — anything from a few million up to roughly 10B parameters. In practice I'd start with Qwen3.5-0.8B or a small Gemma (2B); a 16–24 GB card gives you room for Gemma 9B or Phi-4 14B, and Gemini Nano-1 (1.8B) / Nano-2 (3.25B) are also on the table.&lt;/p&gt;

&lt;h2&gt;
  
  
  Requirements before you touch any code
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;One consumer GPU with at least 8 GB of VRAM. Phi-4 14B will want a lot more than that.&lt;/li&gt;
&lt;li&gt;Python 3.10+.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;transformers&lt;/code&gt;, &lt;code&gt;peft&lt;/code&gt; and &lt;code&gt;datasets&lt;/code&gt; installed.&lt;/li&gt;
&lt;li&gt;A base model picked by name, so you're not guessing later.&lt;/li&gt;
&lt;li&gt;A &lt;code&gt;train.jsonl&lt;/code&gt; file with a few hundred clean examples.&lt;/li&gt;
&lt;li&gt;Admin rights for CUDA drivers and around 20 GB of free disk for checkpoints.
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip install transformers peft datasets
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's the whole environment for a first run. Resist the urge to add a dozen more packages before you've seen the loss go down once.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prepare the data, then pick the model
&lt;/h2&gt;

&lt;p&gt;I do data first because it's the cheapest thing to get wrong and the most annoying thing to debug mid-training. One JSON object per line, one instruction/output pair each:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"instruction"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Classify the ticket"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"output"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"billing"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"instruction"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Classify the ticket"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"output"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"technical"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The trainer expects plain text pairs, not a database dump. For a narrow task, 100–300 clean examples are usually enough to see a style or format shift — quality beats volume here.&lt;/p&gt;

&lt;p&gt;Model choice is mostly a VRAM decision. For an 8 GB card, Qwen3.5-0.8B or Gemma 2B leaves headroom for gradients and optimizer states. On 16–24 GB you can move up to Gemma 9B or Phi-4 14B. Smaller models are not a downgrade for this workflow; they're the reason it fits at all.&lt;/p&gt;

&lt;h2&gt;
  
  
  Load the base model and attach an adapter
&lt;/h2&gt;

&lt;p&gt;Standard Hugging Face loading, nothing exotic. The model downloads once and is cached locally afterwards.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;transformers&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;AutoModelForCausalLM&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;AutoTokenizer&lt;/span&gt;
&lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;AutoModelForCausalLM&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;from_pretrained&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"Qwen/Qwen3.5-0.8B"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;tok&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;AutoTokenizer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;from_pretrained&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"Qwen/Qwen3.5-0.8B"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then wrap it with &lt;code&gt;peft&lt;/code&gt;. I used a small rank and targeted the attention projections:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;peft&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;LoraConfig&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;get_peft_model&lt;/span&gt;
&lt;span class="n"&gt;cfg&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;LoraConfig&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;lora_alpha&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;16&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;target_modules&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"q_proj"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"v_proj"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;get_peft_model&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;cfg&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;print_trainable_parameters&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;print_trainable_parameters()&lt;/code&gt; is the sanity check I now run every single time. You want to see something in the neighbourhood of "0.5% of parameters trainable" and a non-zero trainable count. If it says zero, your &lt;code&gt;target_modules&lt;/code&gt; don't match the architecture and nothing useful will happen during training.&lt;/p&gt;

&lt;h2&gt;
  
  
  Train for a couple of epochs, then test
&lt;/h2&gt;

&lt;p&gt;One to three epochs is plenty for a narrow task. I kept the batch size at 1–4 and used gradient accumulation to get an effective batch that fits in memory. What you're watching is the loss: on my run it moved from around 2.5 down toward 1.0. A gentle downward slope is the signal; a flat line is not.&lt;/p&gt;

&lt;p&gt;Once training finishes, save the adapter:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;save_pretrained&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"my-lora-adapter"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then load the base model plus the adapter and ask it something from your own domain. The comparison that matters is behaviour, not loss: do answers follow the style and format of your dataset? If yes, you're done with the first iteration.&lt;/p&gt;

&lt;p&gt;Before you ship a fine-tuned model anywhere, evaluate it properly rather than eyeballing three prompts.&lt;/p&gt;

&lt;h2&gt;
  
  
  When it breaks
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;CUDA out of memory on the first steps.&lt;/strong&gt; Drop to a smaller model (Gemma 9B → Gemma 2B or Qwen3.5-0.8B), set batch size to 1, or lower the LoRA rank from 16 to 8. In that order — the model swap usually fixes it fastest.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Loss stays flat after hundreds of steps.&lt;/strong&gt; Validate the JSONL first: every line must be a complete, valid JSON object with both &lt;code&gt;instruction&lt;/code&gt; and &lt;code&gt;output&lt;/code&gt;. Then nudge the learning rate up slightly and re-check that the adapter targets the right modules.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The fine-tuned model outputs gibberish.&lt;/strong&gt; This is almost always too many epochs on too little data. Retrain for 1–2 epochs. Also confirm that the tokenizer used at inference is the same one used in training — a mismatch produces exactly this kind of broken text.&lt;/p&gt;

&lt;p&gt;A pre-flight list I keep next to me: the model fits in VRAM with headroom, &lt;code&gt;train.jsonl&lt;/code&gt; holds at least 100–300 clean examples, &lt;code&gt;peft&lt;/code&gt; and &lt;code&gt;transformers&lt;/code&gt; are installed, an output directory exists for the adapter, and my test questions were written down before training started so I can't rationalise the results afterwards.&lt;/p&gt;

&lt;h2&gt;
  
  
  A few questions I had
&lt;/h2&gt;

&lt;p&gt;Can Phi-4 14B be trained on one GPU? Yes, but it needs a high-VRAM consumer card. On anything smaller, start with Qwen3.5-0.8B or Gemma 2B.&lt;/p&gt;

&lt;p&gt;Does a fine-tuned SLM replace a large LLM? No. Small models have their own niche — narrow tasks and agentic systems — and a 0.8B model with a LoRA adapter is not a general-purpose replacement.&lt;/p&gt;

&lt;p&gt;How much data do I need? For a narrow task, a few hundred clean JSONL examples are typically enough to see the format change you're after.&lt;/p&gt;

&lt;p&gt;The takeaway: LoRA is less about compute and more about discipline. Pick a model that fits with room to spare, keep the dataset small and clean, watch &lt;code&gt;print_trainable_parameters()&lt;/code&gt; and the loss curve, and train for fewer epochs than you think you need. If you haven't run a model locally at all yet, get that working first — fine-tuning builds directly on it.&lt;/p&gt;

&lt;p&gt;This article was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt;. More guides like this land in the channel: &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>tutorial</category>
      <category>python</category>
    </item>
    <item>
      <title>One Endpoint for Claude, GPT, DeepSeek and GLM: A Router Setup Guide</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Sat, 14 Nov 2026 10:11:58 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/one-endpoint-for-claude-gpt-deepseek-and-glm-a-router-setup-guide-221m</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/one-endpoint-for-claude-gpt-deepseek-and-glm-a-router-setup-guide-221m</guid>
      <description>&lt;p&gt;I have run four vendor SDKs, four API keys and four billing dashboards side by side, and it was never worth the effort. An API router (also called an aggregator or a gateway) gives you one HTTP endpoint and lets you name the model family per request. That is how I reach Claude, GPT, DeepSeek and GLM from a single Python client today.&lt;/p&gt;

&lt;p&gt;The setup below takes about 30 minutes and sits at intermediate level. You need a terminal with &lt;code&gt;curl&lt;/code&gt; and &lt;code&gt;jq&lt;/code&gt; 1.6+, Python 3.10+ with &lt;code&gt;pip install openai&lt;/code&gt;, and an API key from one router. If you go self-hosted you also need vendor keys and a machine where a proxy can listen on port 4000. The Python client stays under 100 MB of disk; the proxy adds a few hundred more.&lt;/p&gt;

&lt;p&gt;One rule before anything else: router names, model IDs, prices and free-tier limits change monthly. Anything below that depends on a specific vendor is a placeholder. Read the real values from &lt;code&gt;/v1/models&lt;/code&gt; and the router's pricing page on the day you set this up, not from any article.&lt;/p&gt;

&lt;h2&gt;
  
  
  What "OpenAI-compatible" actually gets you
&lt;/h2&gt;

&lt;p&gt;A router is one HTTP endpoint that forwards your request to whichever vendor you name, so your code talks to a single URL instead of four SDKs. The "OpenAI-compatible" part is what makes that practical: the endpoint accepts the same request body as OpenAI's &lt;code&gt;/v1/chat/completions&lt;/code&gt;. That is why the Python snippet further down works unchanged against any router that claims compatibility, and why moving from GPT to DeepSeek becomes a string change rather than a refactor.&lt;/p&gt;

&lt;p&gt;All four families I care about speak that shape. Claude goes through an OpenAI-compatible path or a native Anthropic passthrough, GPT, DeepSeek and GLM (Zhipu) go through &lt;code&gt;/v1/chat/completions&lt;/code&gt;. What actually differs is the exact model IDs, the price per million input and output tokens, the context window and, for DeepSeek, whether a free or trial quota applies. Those are the four things to confirm in the vendor docs before you commit to anything.&lt;/p&gt;

&lt;h2&gt;
  
  
  Hosted aggregator or self-hosted proxy?
&lt;/h2&gt;

&lt;p&gt;Pick the shape before you write code, because the credential model is different.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Criterion&lt;/th&gt;
&lt;th&gt;Hosted aggregator&lt;/th&gt;
&lt;th&gt;Self-hosted proxy&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;What it is&lt;/td&gt;
&lt;td&gt;A vendor-run API that already holds accounts with Anthropic, OpenAI, DeepSeek and Zhipu&lt;/td&gt;
&lt;td&gt;A process you run, for example LiteLLM, holding &lt;em&gt;your&lt;/em&gt; vendor keys and exposing one local port&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Credentials you manage&lt;/td&gt;
&lt;td&gt;One router key&lt;/td&gt;
&lt;td&gt;Four vendor keys plus a local proxy key&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Free/cheap access&lt;/td&gt;
&lt;td&gt;Depends on current trial credits and per-model pricing, verify on the pricing page&lt;/td&gt;
&lt;td&gt;You pay each vendor at list price, no markup and no free tier added&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Billing&lt;/td&gt;
&lt;td&gt;One invoice, vendor-level detail varies&lt;/td&gt;
&lt;td&gt;Per-vendor invoices with full detail&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Data path&lt;/td&gt;
&lt;td&gt;Prompts pass through a third party, so check its retention policy&lt;/td&gt;
&lt;td&gt;Prompts go only to the vendors you configured&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Best fit&lt;/td&gt;
&lt;td&gt;Fast prototyping, one-key demos&lt;/td&gt;
&lt;td&gt;Teams with existing vendor contracts or compliance constraints&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;In practice I start hosted when I want one key and no YAML, and move to the proxy when a prompt cannot leave my own infrastructure.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 1: collect the base URL, the key and the model IDs
&lt;/h2&gt;

&lt;p&gt;Three strings drive everything after this: base URL, key, model ID. Guessing the model ID is the most common reason a &lt;code&gt;404&lt;/code&gt; shows up two steps later, so read the list from the API instead of copying it from a blog post.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="kd"&gt;export&lt;/span&gt; &lt;span class="nv"&gt;ROUTER_BASE_URL&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"https://&amp;lt;router-base-url&amp;gt;/v1"&lt;/span&gt;
&lt;span class="kd"&gt;export&lt;/span&gt; &lt;span class="nv"&gt;ROUTER_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"&amp;lt;your-api-key&amp;gt;"&lt;/span&gt;

curl -s &lt;span class="s2"&gt;"$ROUTER_BASE_URL/models"&lt;/span&gt; \
  -H &lt;span class="s2"&gt;"Authorization: Bearer $ROUTER_API_KEY"&lt;/span&gt; \
  &lt;span class="o"&gt;|&lt;/span&gt; jq -r &lt;span class="s1"&gt;'.data[].id'&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; sort &lt;span class="o"&gt;|&lt;/span&gt; head -20
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Illustrative output, where your IDs will differ:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;anthropic/&amp;lt;claude-model-id&amp;gt;
deepseek/&amp;lt;deepseek-model-id&amp;gt;
openai/&amp;lt;gpt-model-id&amp;gt;
zhipu/&amp;lt;glm-model-id&amp;gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This works because &lt;code&gt;/v1/models&lt;/code&gt; is part of the compatible surface, so a router that claims compatibility has to answer it. If HTML comes back instead of JSON, your base URL is wrong. Once you have the four IDs, pin them in a local file so later steps stay reproducible:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;cat &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; models.env &lt;span class="o"&gt;&amp;lt;&amp;lt;&lt;/span&gt;&lt;span class="s1"&gt;'EOF'&lt;/span&gt;
&lt;span class="nv"&gt;CLAUDE_MODEL&lt;/span&gt;&lt;span class="o"&gt;=&amp;lt;&lt;/span&gt;id-from-the-list-above&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;
&lt;span class="nv"&gt;GPT_MODEL&lt;/span&gt;&lt;span class="o"&gt;=&amp;lt;&lt;/span&gt;id-from-the-list-above&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;
&lt;span class="nv"&gt;DEEPSEEK_MODEL&lt;/span&gt;&lt;span class="o"&gt;=&amp;lt;&lt;/span&gt;id-from-the-list-above&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;
&lt;span class="nv"&gt;GLM_MODEL&lt;/span&gt;&lt;span class="o"&gt;=&amp;lt;&lt;/span&gt;id-from-the-list-above&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;
EOF
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Check it with &lt;code&gt;source models.env &amp;amp;&amp;amp; echo $CLAUDE_MODEL&lt;/code&gt;, which should print an ID and not an empty line.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: point one client at the router
&lt;/h2&gt;

&lt;p&gt;Overriding &lt;code&gt;base_url&lt;/code&gt; in the OpenAI SDK replaces four vendor SDKs. Model choice becomes a dictionary lookup, and a fallback becomes a retry.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;
&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"ROUTER_BASE_URL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"ROUTER_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;timeout&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mf"&gt;60.0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="no"&gt;MODELS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="s2"&gt;"claude"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"CLAUDE_MODEL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"gpt"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"GPT_MODEL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"deepseek"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"DEEPSEEK_MODEL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"glm"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"GLM_MODEL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;ask&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;family&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;str&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;prompt&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;str&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="kt"&gt;str&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completions&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="no"&gt;MODELS&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;family&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"role"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"user"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"content"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;prompt&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;max_tokens&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;128&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;family&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;usage&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;prompt_tokens&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;usage&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completion_tokens&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;choices&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;message&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;content&lt;/span&gt;

&lt;span class="k"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;ask&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"deepseek"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"Reply with the single word: ready"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A run prints something like &lt;code&gt;deepseek 12 1&lt;/code&gt; and then &lt;code&gt;ready&lt;/code&gt;. Keep those token counts: multiplied by the per-million prices from Step 1, they give you the real cost of the call, which is more honest than a headline price.&lt;/p&gt;

&lt;p&gt;For the self-hosted route, all credentials stay on the server and clients get only a local key:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;model_list&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
  &lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="na"&gt;model_name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;claude&lt;/span&gt;
    &lt;span class="na"&gt;litellm_params&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;anthropic/&amp;lt;anthropic-model-id&amp;gt;&lt;/span&gt;
      &lt;span class="na"&gt;api_key&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;os.environ/ANTHROPIC_API_KEY&lt;/span&gt;
  &lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="na"&gt;model_name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;gpt&lt;/span&gt;
    &lt;span class="na"&gt;litellm_params&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;openai/&amp;lt;openai-model-id&amp;gt;&lt;/span&gt;
      &lt;span class="na"&gt;api_key&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;os.environ/OPENAI_API_KEY&lt;/span&gt;
  &lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="na"&gt;model_name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;deepseek&lt;/span&gt;
    &lt;span class="na"&gt;litellm_params&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;deepseek/&amp;lt;deepseek-model-id&amp;gt;&lt;/span&gt;
      &lt;span class="na"&gt;api_key&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;os.environ/DEEPSEEK_API_KEY&lt;/span&gt;
  &lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="na"&gt;model_name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;glm&lt;/span&gt;
    &lt;span class="na"&gt;litellm_params&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;&amp;lt;provider-prefix&amp;gt;/&amp;lt;glm-model-id&amp;gt;&lt;/span&gt;
      &lt;span class="na"&gt;api_key&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;os.environ/ZHIPU_API_KEY&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;litellm --config config.yaml --port 4000
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The proxy translates the OpenAI schema into each vendor's API and maps &lt;code&gt;model_name&lt;/code&gt; onto a real model ID, so your app only ever sends &lt;code&gt;claude&lt;/code&gt; or &lt;code&gt;glm&lt;/code&gt;. The GLM provider prefix and the exact CLI flag are version-specific, so check the proxy docs if startup fails. Verify with &lt;code&gt;curl -s localhost:4000/v1/models -H "Authorization: Bearer &amp;lt;local-key&amp;gt;" | jq '.data | length'&lt;/code&gt;, which should return &lt;code&gt;4&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: smoke test answers, latency and cost
&lt;/h2&gt;

&lt;p&gt;A router can list a model and still fail on the call, because of a provider outage, an exhausted quota or a permission you are missing. I would rather find that in a test than in a demo, so I loop over all four families and capture status, timing and the usage block in one pass:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="k"&gt;for&lt;/span&gt; m &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="s2"&gt;"$CLAUDE_MODEL"&lt;/span&gt; &lt;span class="s2"&gt;"$GPT_MODEL"&lt;/span&gt; &lt;span class="s2"&gt;"$DEEPSEEK_MODEL"&lt;/span&gt; &lt;span class="s2"&gt;"$GLM_MODEL"&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt; &lt;span class="k"&gt;do&lt;/span&gt;
  curl -s -o /tmp/out.json -w &lt;span class="s2"&gt;"%{http_code} %{time_total}s\n"&lt;/span&gt; \
    &lt;span class="s2"&gt;"$ROUTER_BASE_URL/chat/completions"&lt;/span&gt; \
    -H &lt;span class="s2"&gt;"Authorization: Bearer $ROUTER_API_KEY"&lt;/span&gt; \
    -H &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; \
    -d &lt;span class="s2"&gt;"{\"model\":\"$m\",\"messages\":[{\"role\":\"user\",\"content\":\"Reply with the single word: ready\"}],\"max_tokens\":16}"&lt;/span&gt;
  jq -r &lt;span class="s1"&gt;'.usage // "no usage block"'&lt;/span&gt; /tmp/out.json
&lt;span class="k"&gt;done&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Two things to look at. If a model answers but returns no &lt;code&gt;usage&lt;/code&gt; block, you cannot compute cost from the response, and you should treat that as a limitation before building anything on top of it. And round-trip time in the &lt;code&gt;time_total&lt;/code&gt; column is what you compare between families when you pick a default.&lt;/p&gt;

&lt;p&gt;Finally, break a key on purpose: swap one character in &lt;code&gt;ROUTER_API_KEY&lt;/code&gt; and run a single call. The error shape tells you how the router reports auth failures and quota exhaustion, which is exactly what you need to write a retry or fallback later.&lt;/p&gt;

&lt;h2&gt;
  
  
  The six-point checklist before you spend money
&lt;/h2&gt;

&lt;p&gt;I tick all of these on the router's docs page before putting a card down:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;[ ] The model list contains all four families you actually need&lt;/li&gt;
&lt;li&gt;[ ] Free trial or free-tier conditions are stated explicitly, not implied&lt;/li&gt;
&lt;li&gt;[ ] Price per million input and output tokens is listed per model&lt;/li&gt;
&lt;li&gt;[ ] Rate limits (requests per minute, tokens per minute) are documented&lt;/li&gt;
&lt;li&gt;[ ] Prompt logging and retention policy is written down&lt;/li&gt;
&lt;li&gt;[ ] The OpenAI-compatible base path is documented as &lt;code&gt;/v1&lt;/code&gt;, not &lt;code&gt;/api&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If two or more boxes are empty, keep looking. The cheap part of a router is the price; the expensive part is discovering the gaps after your code depends on it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;A router turns four integrations into one base URL and a small dictionary of model IDs. Hosted gets you there fastest and exposes you to a third party's retention policy; self-hosted keeps prompts in your own path but means managing four sets of vendor keys. Either way, the discipline is the same: pull IDs from &lt;code&gt;/v1/models&lt;/code&gt;, record the usage numbers on a real call, and verify the pricing and rate limits yourself, because all of it will be different by the time you read this.&lt;/p&gt;

&lt;p&gt;This article was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt;. Shorter notes and updates on the same topic end up in the Telegram channel: &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>tutorial</category>
      <category>programming</category>
    </item>
    <item>
      <title>Turning the Microsoft Store's 49 Free Apps Into a Verified AI Shortlist</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Sat, 14 Nov 2026 10:05:57 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/turning-the-microsoft-stores-49-free-apps-into-a-verified-ai-shortlist-4kl5</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/turning-the-microsoft-stores-49-free-apps-into-a-verified-ai-shortlist-4kl5</guid>
      <description>&lt;p&gt;I kept seeing the same claim: the Microsoft Store has a "Best free apps" collection with 49 free Windows apps. What nobody posts is the actual list — the selection criteria aren't disclosed and not a single app name is published. Rather than guess, I spent 25 minutes turning that page into something verifiable on my own machine: a CSV shortlist with app name, publisher, price field, install source and a pass/fail note for each candidate.&lt;/p&gt;

&lt;p&gt;Here's the routine. Basic difficulty, and everything you need already ships with a Windows PC.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you need first
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Windows PC with a working Microsoft Store&lt;/li&gt;
&lt;li&gt;Microsoft account signed in to the Store (free downloads still go through the account)&lt;/li&gt;
&lt;li&gt;Local administrator rights&lt;/li&gt;
&lt;li&gt;Enough free disk space — the collection page doesn't list sizes, so check each app's own size line&lt;/li&gt;
&lt;li&gt;PowerShell, which is built into Windows&lt;/li&gt;
&lt;li&gt;Optional: &lt;code&gt;winget&lt;/code&gt;, for scriptable search and install&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Two checks up front save most of the pain later. First, record your build:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="n"&gt;winver&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You'll get a dialog with edition and build, e.g. &lt;code&gt;Version 23H2 (OS Build 22631.xxxx)&lt;/code&gt;. Every Store listing carries its own minimum requirements; if your build is lower, read that app's documentation before installing.&lt;/p&gt;

&lt;p&gt;Second, confirm the Store package itself is healthy:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Get&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;AppxPackage&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Name&lt;/span&gt; &lt;span class="nc"&gt;Microsoft&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;WindowsStore&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Select&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="nc"&gt;Name&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;Version&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If the version column is empty or nothing comes back, the Store package is missing or damaged — fix that before browsing anything.&lt;/p&gt;

&lt;p&gt;Then take a baseline snapshot of what is already installed:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Get&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;AppxPackage&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Where&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="nc"&gt;SignatureKind&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;eq&lt;/span&gt; &lt;span class="s1"&gt;'Store'&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt;
  &lt;span class="nc"&gt;Sort&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="nc"&gt;Name&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Select&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="nc"&gt;Name&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;Version&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt;
  &lt;span class="nc"&gt;Export&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Csv&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Path&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Desktop\store-baseline.csv"&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;NoTypeInformation&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Encoding&lt;/span&gt; &lt;span class="no"&gt;UTF8&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You now have a "before" file on your Desktop. Without it, a leftover app later looks like something the collection added.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 1: Open the collection and count what you see
&lt;/h2&gt;

&lt;p&gt;Open the Store app and find the collection by name inside the app — the posts don't publish a URL, so don't go typing addresses into a browser. The page should show 49 entries. If it shows a different number, the collection changed since the posts were written: write down what you see and move on. Don't force the figure.&lt;/p&gt;

&lt;p&gt;Start your shortlist file as you go:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nv"&gt;$rows&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; @&lt;span class="p"&gt;(&lt;/span&gt;
  &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="kt"&gt;pscustomobject&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;@&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nc"&gt;App&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s1"&gt;'&amp;lt;name on the page&amp;gt;'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nc"&gt;Publisher&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s1"&gt;'&amp;lt;publisher&amp;gt;'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nc"&gt;Price&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s1"&gt;'Free'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nc"&gt;Category&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s1"&gt;'&amp;lt;category&amp;gt;'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nc"&gt;Installed&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s1"&gt;'no'&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="nv"&gt;$rows&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Export&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Csv&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Path&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Desktop\store-shortlist.csv"&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;NoTypeInformation&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Encoding&lt;/span&gt; &lt;span class="no"&gt;UTF8&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For each candidate, click it, copy the name and publisher, add one line and re-run the block.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: Find AI entries and read each listing before installing
&lt;/h2&gt;

&lt;p&gt;Since no AI app is named in the source posts, treat every candidate as unverified until you've read its own page. &lt;code&gt;winget&lt;/code&gt; is the fastest way to scan:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="n"&gt;winget&lt;/span&gt; &lt;span class="n"&gt;search&lt;/span&gt; &lt;span class="o"&gt;--&lt;/span&gt;&lt;span class="n"&gt;source&lt;/span&gt; &lt;span class="n"&gt;msstore&lt;/span&gt; &lt;span class="s2"&gt;"AI"&lt;/span&gt;
&lt;span class="n"&gt;winget&lt;/span&gt; &lt;span class="n"&gt;search&lt;/span&gt; &lt;span class="o"&gt;--&lt;/span&gt;&lt;span class="n"&gt;source&lt;/span&gt; &lt;span class="n"&gt;msstore&lt;/span&gt; &lt;span class="s2"&gt;"assistant"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You get Name, Id, Version and Source. The Id is what you reuse to install:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="n"&gt;winget&lt;/span&gt; &lt;span class="n"&gt;install&lt;/span&gt; &lt;span class="o"&gt;--&lt;/span&gt;&lt;span class="n"&gt;source&lt;/span&gt; &lt;span class="n"&gt;msstore&lt;/span&gt; &lt;span class="o"&gt;--&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="o"&gt;--&lt;/span&gt;&lt;span class="n"&gt;accept&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;package&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;agreements&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Before you run that, check five things on the Store page:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Publisher is a company you can actually identify, not a lookalike name&lt;/li&gt;
&lt;li&gt;Price reads &lt;code&gt;Free&lt;/code&gt;, and the feature you want isn't behind a subscription — the collection page's own descriptions mention Microsoft 365, Office and Azure&lt;/li&gt;
&lt;li&gt;Permissions match what the app claims to do&lt;/li&gt;
&lt;li&gt;The size line fits your free disk space&lt;/li&gt;
&lt;li&gt;The app shows up in the Store app's search, not just in a browser&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;"Free" in a listing means the download is free. That's it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: Prove each app works, then finish the shortlist
&lt;/h2&gt;

&lt;p&gt;An icon in the Start menu proves nothing — sign-in, model downloads and permissions all tend to fail after installation. Launch each app, run one real task in it, then export the "after" state and compare it with the baseline:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Get&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;AppxPackage&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Where&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="nc"&gt;SignatureKind&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;eq&lt;/span&gt; &lt;span class="s1"&gt;'Store'&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt;
  &lt;span class="nc"&gt;Sort&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="nc"&gt;Name&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Select&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="nc"&gt;Name&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;Version&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt;
  &lt;span class="nc"&gt;Export&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Csv&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Path&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Desktop\store-after.csv"&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;NoTypeInformation&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Encoding&lt;/span&gt; &lt;span class="no"&gt;UTF8&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Diff the two Name columns: whatever is new is what you actually added, including anything you installed by accident while clicking around. Then fill in &lt;code&gt;Installed&lt;/code&gt; and a pass/fail note for every row of &lt;code&gt;store-shortlist.csv&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three things the source posts actually name
&lt;/h2&gt;

&lt;p&gt;Worth being explicit, because it explains why the shortlist has to be built by hand:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Entry&lt;/th&gt;
&lt;th&gt;What's confirmed&lt;/th&gt;
&lt;th&gt;Price&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Microsoft Store "Best free apps" collection&lt;/td&gt;
&lt;td&gt;One page, 49 free Windows apps; descriptions mention Microsoft 365, Office, Azure; criteria not disclosed; no app names published&lt;/td&gt;
&lt;td&gt;Free apps&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini 4 Argon&lt;/td&gt;
&lt;td&gt;A new Google model discussed on Hacker News — 1529 points, 1011 comments; not stated to be in the Store&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;MLBB-waydroid&lt;/td&gt;
&lt;td&gt;Runs Mobile Legends: Bang Bang on Linux via Waydroid and streams it to an Android phone with Sunshine; not a Store app in the posts&lt;/td&gt;
&lt;td&gt;Not stated&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Those are the only items named anywhere in the source. Two of the three aren't even Store apps — so if a listicle promises you the best free AI apps in the Microsoft Store and names nothing, this is why.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;The value here isn't a list someone hands you — it's the shortlist you can defend. Record your build, snapshot the baseline, count the collection yourself, read each listing's publisher, price, permissions and size, then diff the after-state and write pass/fail into the CSV. That's about 25 minutes for a file that's true on your machine, which is more than any unnamed collection page can give you.&lt;/p&gt;

&lt;p&gt;This article was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt;. Shorter notes and new guides go out on the channel: &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>windows</category>
      <category>ai</category>
      <category>tutorial</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Picking an AI app for Android in 2026: what actually matters</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Fri, 13 Nov 2026 10:11:25 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/picking-an-ai-app-for-android-in-2026-what-actually-matters-59im</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/picking-an-ai-app-for-android-in-2026-what-actually-matters-59im</guid>
      <description>&lt;p&gt;AI has moved from a novelty on Android to something closer to a built-in layer of the OS. In 2026 you can draft a message, compress a long article into bullets, or generate an image without ever opening a laptop. The hard part is no longer finding an AI app — it is picking the one that matches how you actually use your phone. Here is how I would break the landscape down.&lt;/p&gt;

&lt;h2&gt;
  
  
  The phone became the default place people use AI
&lt;/h2&gt;

&lt;p&gt;Android is the most widely used mobile operating system, and AI is now its most important feature layer. These apps are no longer experimental: they handle everyday work like drafting messages, summarizing long reads, and creating images on the go.&lt;/p&gt;

&lt;p&gt;The convenience is the whole point. A good app has access to your microphone and camera, works in the flow of whatever you are doing, and can replace a couple of separate tools at once. For a lot of people, the smartphone is now the primary device for interacting with AI — not the desktop.&lt;/p&gt;

&lt;h2&gt;
  
  
  Chat assistants: pick the ecosystem, not the feature list
&lt;/h2&gt;

&lt;p&gt;The all-rounders dominate. ChatGPT, Google Gemini, and Microsoft Copilot cover text generation, question answering, and basic reasoning right on the phone, and they are the category most people start with.&lt;/p&gt;

&lt;p&gt;They are best for quick answers, brainstorming, and rewriting text. In 2026 the gap between them has narrowed enough that the decision usually comes down to what you already use. Live inside Google services? Gemini slots in naturally. Prefer OpenAI models? ChatGPT stays a solid pick.&lt;/p&gt;

&lt;p&gt;One thing worth checking: most of these support voice input, which makes them genuinely useful hands-free. Some also let you share your screen or upload a photo, which turns the assistant into a visual helper — reading a label or explaining a diagram, for example.&lt;/p&gt;

&lt;h2&gt;
  
  
  Writing tools: offline support is the feature to watch
&lt;/h2&gt;

&lt;p&gt;For people who draft emails, posts, or documents on a phone, specialized writing apps are the better fit than a general chatbot. They focus on tone adjustment, grammar correction, and text expansion.&lt;/p&gt;

&lt;p&gt;The best writing tools in 2026 do the basic work on-device and only reach for the cloud when the task is heavy. That combination matters twice: it helps with privacy, and it helps if you are on a limited data plan. Many productivity apps now bake AI straight into the keyboard, so you can rewrite a sentence without jumping between apps.&lt;/p&gt;

&lt;p&gt;Summarization is the other big win, especially for students and professionals. Squeezing a long PDF or web article into a handful of bullet points is a real time-saver when the screen is small.&lt;/p&gt;

&lt;h2&gt;
  
  
  Image apps: fastest-growing category, most caveats
&lt;/h2&gt;

&lt;p&gt;Image generation and editing is the quickest-moving space on Android right now. These apps turn text prompts into pictures, remove objects from photos, and upscale old images.&lt;/p&gt;

&lt;p&gt;Mobile image models in 2026 are good enough for social media content and personal projects. What makes a good one is boring and practical: a simple prompt box, a gallery of styles, and one-tap export. Batch processing is a nice bonus when you want several variations of the same idea.&lt;/p&gt;

&lt;p&gt;The catch is cost. High-quality generation usually needs a paid subscription, and free tiers tend to add watermarks or cap how many images you get per day.&lt;/p&gt;

&lt;h2&gt;
  
  
  A short pre-install checklist
&lt;/h2&gt;

&lt;p&gt;Before you commit storage and permissions to an app, run through this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;1. Main need      -&amp;gt; chat / writing / images
2. Privacy policy -&amp;gt; what is collected, is it used for training
3. Data usage     -&amp;gt; does it work offline for simple tasks
4. Updates        -&amp;gt; how often does it ship
5. Reviews        -&amp;gt; read the recent ones, not the top ones
6. Footprint      -&amp;gt; storage + battery, is there a lightweight mode
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;App updates are an underrated signal: apps that ship regularly tend to improve faster and patch security issues sooner. And read recent user reviews rather than the all-time best ones — the AI landscape moves quickly, and last year's favorite may already be behind.&lt;/p&gt;

&lt;p&gt;On hardware: some AI apps are heavy. Look for lightweight modes or cloud processing options so your phone stays responsive.&lt;/p&gt;

&lt;h2&gt;
  
  
  Privacy is where the real differences show up
&lt;/h2&gt;

&lt;p&gt;AI apps touch your text, voice, and images, so it is fair to ask what happens to them. In 2026 more apps offer on-device processing for sensitive tasks, which means the data never leaves the phone. For business users or anyone handling personal information, that is a meaningful advantage.&lt;/p&gt;

&lt;p&gt;Two things I check every time: what the app collects, and whether it is used for training. Reputable apps give you a way to opt out. Be wary of anything asking for full photo library or contacts access when its core feature does not need it.&lt;/p&gt;

&lt;p&gt;One quick note before you go: I also keep a shorter version of these app roundups over at AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt; — and the channel at &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt; is where the daily links land.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;There is no single best AI app for Android in 2026, and anyone claiming otherwise is probably selling something. Start from your main task: a chat assistant with a strong free tier if you want a general helper, a writing tool with offline mode if you type a lot, an image app with good style controls if you make visuals. Then filter by privacy policy, update frequency, and how much battery the thing eats. Do that, and you will end up with two or three apps that actually earn their home screen space.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>android</category>
      <category>productivity</category>
      <category>beginners</category>
    </item>
    <item>
      <title>Testing offline Android translators with adb and airplane mode</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Fri, 13 Nov 2026 10:05:25 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/testing-offline-android-translators-with-adb-and-airplane-mode-22ll</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/testing-offline-android-translators-with-adb-and-airplane-mode-22ll</guid>
      <description>&lt;p&gt;Offline translation is one of those claims that's easy to make and annoying to verify. You tap a button, text appears, and you have no idea whether the phone quietly grabbed a mobile data connection behind your back. So I set up a small, repeatable test: same phone, same sentence, same target language, airplane mode on, and &lt;code&gt;adb&lt;/code&gt; capturing the evidence as files instead of my memory of what I saw on screen.&lt;/p&gt;

&lt;p&gt;The two contenders were Google Translate (installed from Google Play) and InkPeek, a translator that runs an AI model locally on the device. The whole run takes about 25 minutes and sits at intermediate level, mostly because you drive the phone from a PC.&lt;/p&gt;

&lt;p&gt;One warning before we start: my source material doesn't state InkPeek's package name, minimum Android version, model size or price. Wherever those matter below I've marked them as things to confirm in the vendor's own documentation before you commit to anything.&lt;/p&gt;

&lt;h2&gt;
  
  
  The ground rules I used
&lt;/h2&gt;

&lt;p&gt;A fair comparison is mostly about removing variables:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Airplane mode on, Wi-Fi and mobile data both disabled.&lt;/li&gt;
&lt;li&gt;The same source and target language pair for both apps.&lt;/li&gt;
&lt;li&gt;The same sentence, e.g. &lt;code&gt;The meeting is moved to Thursday at 9 a.m.&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Results saved as screenshots on the PC via &lt;code&gt;adb&lt;/code&gt;, not judged from the screen.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The pass condition: the translation is readable, the meaning holds, and the app shows no network error such as a "connect to the internet" banner.&lt;/p&gt;

&lt;p&gt;For hardware I used a USB cable, a PC with Android Platform Tools, and a phone on Android 11 or newer (the &lt;code&gt;svc&lt;/code&gt; commands below are friendlier there). Free storage matters too — a few hundred megabytes at minimum, since every language pack or on-device model lives on the phone.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 1: confirm adb can actually talk to the phone
&lt;/h2&gt;

&lt;p&gt;Enable &lt;code&gt;Settings → About phone → tap Build number 7 times&lt;/code&gt;, then flip &lt;code&gt;Settings → System → Developer options → USB debugging&lt;/code&gt; on. Plug in the cable, accept &lt;code&gt;Allow USB debugging?&lt;/code&gt; and tick &lt;code&gt;Always allow from this computer&lt;/code&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb devices
adb shell getprop ro.build.version.release
adb shell getprop ro.product.model
adb shell df -h /data
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;On a working setup the first line comes back as a serial followed by the word &lt;code&gt;device&lt;/code&gt; — not &lt;code&gt;unauthorized&lt;/code&gt;, not &lt;code&gt;offline&lt;/code&gt;. In my run the phone reported &lt;code&gt;Pixel 7&lt;/code&gt; and &lt;code&gt;/data&lt;/code&gt; showed &lt;code&gt;46G&lt;/code&gt; available, plenty of headroom for several packs.&lt;/p&gt;

&lt;p&gt;If you see &lt;code&gt;unauthorized&lt;/code&gt;, nothing further in this guide will work. Unlock the phone, accept the prompt again, and re-run &lt;code&gt;adb devices&lt;/code&gt; until the serial shows &lt;code&gt;device&lt;/code&gt; twice in a row.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: put both apps into an offline-ready state
&lt;/h2&gt;

&lt;p&gt;This is the step people skip, and it's the one that decides whether the test means anything. An offline translator is only offline for the languages it has already stored locally.&lt;/p&gt;

&lt;p&gt;For Google Translate, nothing is installed by default. Open the app, tap the profile icon in the top right, go to &lt;strong&gt;Offline translation&lt;/strong&gt;, pick a language and download it. Repeat for every language in your pair — an English↔Spanish test needs two packs, because each side of the pair is handled locally.&lt;/p&gt;

&lt;p&gt;For InkPeek, install the build from its official distribution channel and let the app pull its own model through its on-screen flow. If you're sideloading instead:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb install -r ~/apk/inkpeek/&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;your-build&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;.apk
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;Success&lt;/code&gt; is what you want. Anything else prints a reason code, for example &lt;code&gt;Failure [INSTALL_FAILED_INSUFFICIENT_STORAGE]&lt;/code&gt;, which usually means the storage check above was optimistic.&lt;/p&gt;

&lt;p&gt;Then verify both packages exist (swap in the real vendor package name for InkPeek):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb shell pm list packages &lt;span class="o"&gt;|&lt;/span&gt; grep -i -E &lt;span class="s2"&gt;"translate|inkpeek"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;package:com.google.android.apps.translate
package:&amp;lt;vendor.package.for.inkpeek&amp;gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I also recorded the Google Translate build so I knew exactly what I'd tested — &lt;code&gt;adb shell dumpsys package com.google.android.apps.translate | grep -m1 versionName&lt;/code&gt; prints &lt;code&gt;versionName=&amp;lt;x.y.z&amp;gt;&lt;/code&gt;. Don't compare that number to one from an article; check the current Play listing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: cut the network and grab proof
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb shell svc wifi disable
adb shell svc data disable
adb shell settings get global airplane_mode_on
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That last command returns &lt;code&gt;1&lt;/code&gt; when airplane mode is on and &lt;code&gt;0&lt;/code&gt; when it isn't. If you get &lt;code&gt;0&lt;/code&gt;, pull down the quick settings shade and enable it before continuing. Leaving one radio alive is the single most common reason a "failed" offline test turns out to have been online the whole time.&lt;/p&gt;

&lt;p&gt;With everything dark, translate the fixed sentence in both apps and capture each screen:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb exec-out screencap -p &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; ~/offline-google.png
adb exec-out screencap -p &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; ~/offline-inkpeek.png
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When you're done, restore connectivity so you're not leaving the phone crippled:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb shell svc wifi &lt;span class="nb"&gt;enable&lt;/span&gt;
adb shell svc data &lt;span class="nb"&gt;enable&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Step 4: judge the files, not your impression
&lt;/h2&gt;

&lt;p&gt;Open both PNGs and check three things: the sentence is readable, the meaning matches the source, and no network-related error is anywhere on screen. Writing the verdict into a table keeps the decision honest:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;App&lt;/th&gt;
&lt;th&gt;Works offline&lt;/th&gt;
&lt;th&gt;Prerequisite&lt;/th&gt;
&lt;th&gt;Price&lt;/th&gt;
&lt;th&gt;Confirm yourself&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Google Translate&lt;/td&gt;
&lt;td&gt;Yes, once packs are downloaded&lt;/td&gt;
&lt;td&gt;One pack per language, in-app&lt;/td&gt;
&lt;td&gt;Free — check the Play listing&lt;/td&gt;
&lt;td&gt;Current listing, pack sizes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;InkPeek&lt;/td&gt;
&lt;td&gt;Yes, with its on-device model installed&lt;/td&gt;
&lt;td&gt;The app's own model files&lt;/td&gt;
&lt;td&gt;Not stated in my source material&lt;/td&gt;
&lt;td&gt;Package name, minimum Android version, model size, licence&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The rule I ended up with: if your everyday language pair already has a downloaded pack and you mostly need short, predictable sentences, Google Translate is the zero-cost default. If you specifically want on-device AI behaviour that the Google app doesn't expose, InkPeek is the one to keep — but only after its model has been downloaded once and the airplane-mode test passes cleanly.&lt;/p&gt;

&lt;h2&gt;
  
  
  Checklist before you trust a result
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;[ ] &lt;code&gt;adb devices&lt;/code&gt; reports &lt;code&gt;device&lt;/code&gt; for your serial&lt;/li&gt;
&lt;li&gt;[ ] Google Translate lists every language of your pair under &lt;strong&gt;Offline translation&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;[ ] InkPeek shows its model as installed rather than offering a download&lt;/li&gt;
&lt;li&gt;[ ] &lt;code&gt;adb shell settings get global airplane_mode_on&lt;/code&gt; prints &lt;code&gt;1&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;[ ] Two screenshots saved to the PC, neither showing a network error&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;The interesting part of this comparison isn't which app wins — it's that "offline" is a property of the installation, not of the app icon. Google Translate needs a pack per language and gives you nothing until you download them; InkPeek needs its model fetched once. After that, both are genuinely usable with the radios dead. The habit worth keeping is the setup: check the device with &lt;code&gt;adb devices&lt;/code&gt;, download the local data first, flip airplane mode, save screenshots, and compare files instead of vibes.&lt;/p&gt;

&lt;p&gt;If you want more write-ups like this one, it was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt; — and the channel posts similar guides at &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>android</category>
      <category>tutorial</category>
      <category>ai</category>
      <category>productivity</category>
    </item>
    <item>
      <title>MacBook Air M5 for AI work: a one-hour buying checklist</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Thu, 12 Nov 2026 10:17:51 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/macbook-air-m5-for-ai-work-a-one-hour-buying-checklist-8po</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/macbook-air-m5-for-ai-work-a-one-hour-buying-checklist-8po</guid>
      <description>&lt;p&gt;I went looking for a spec sheet for the MacBook Air M5 and came back empty-handed. The digest I was working from — an AI news roundup dated 10 October 2026 — has no Apple hardware in it at all: no prices, no memory configurations, no benchmarks. Any hard number you have seen quoted elsewhere should be treated as unverified until you read it on the vendor's own product page.&lt;/p&gt;

&lt;p&gt;So rather than invent figures, here is what I can honestly hand over: a procedure. Roughly an hour of checking that tells you whether an Air-class Mac fits &lt;em&gt;your&lt;/em&gt; AI work. The call comes down to two questions — do your tools run on macOS at all, and do your models fit in the unified memory you are paying for?&lt;/p&gt;

&lt;h2&gt;
  
  
  What the source posts actually contain
&lt;/h2&gt;

&lt;p&gt;Being straight about the evidence matters here. The digest mentions three things: Kommersant's AI coverage (news plus expert commentary), a tool called VRization (desktop-to-phone VR over USB or LAN, with GPU capture and gyroscope control, listed for Windows, Android and iOS), and an offhand reference to old Windows versions. That is it. Nothing about Apple laptops.&lt;/p&gt;

&lt;p&gt;Which means a spec comparison would be fiction. What is not fiction is the habit that protects your wallet: when a forum comment tells you "the M5 Air runs every local model", open the vendor's product page, look for the memory options, and check whether an exact figure is actually listed. If it is not, mark the claim unverified and move on.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prep list, about 20 minutes
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;[ ] Write down every AI tool you use at least once a week — chat app, coding assistant, image generator, local model runner.&lt;/li&gt;
&lt;li&gt;[ ] Open the vendor's system requirements page for each one in its own browser tab.&lt;/li&gt;
&lt;li&gt;[ ] Line up a Mac you can borrow for an hour: a store demo unit, a colleague's laptop, a work machine.&lt;/li&gt;
&lt;li&gt;[ ] Write one test task down &lt;em&gt;before&lt;/em&gt; you start. Something like "summarise a 20-page PDF and answer three questions about it".&lt;/li&gt;
&lt;li&gt;[ ] Turn your budget into the &lt;strong&gt;final price of the exact configuration&lt;/strong&gt;, not the advertised base price.&lt;/li&gt;
&lt;li&gt;[ ] Make sure you have free-tier accounts for the cloud tools you already use — not everything has to run locally.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Six steps to check a machine against your own work
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;1. Confirm macOS support for every tool.&lt;/strong&gt; Search the vendor's download or requirements page for the word "macOS". In our source, VRization lists Windows, Android and iOS — no macOS entry. You want a two-column result: runs on macOS, does not. The OS decides whether a tool runs at all, long before speed enters the picture.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Find out whether anything needs CUDA.&lt;/strong&gt; CUDA is a GPU programming layer tied to NVIDIA hardware. If a tool you rely on requires it, an Apple laptop is the wrong purchase regardless of chip speed. Sort those tools out now, not after the invoice arrives.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. List the local models you plan to run.&lt;/strong&gt; Plain text, one per line, placeholders until you confirm real names:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;# my_local_models.txt — one model per line
&amp;lt;model-A&amp;gt;
&amp;lt;model-B&amp;gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Local models load into unified memory — one pool of RAM shared by the CPU and GPU — so the combined size of the models you want is the number that decides whether a machine is big enough.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Measure the machine you are testing.&lt;/strong&gt; In Terminal:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Which chip, and how much unified memory?&lt;/span&gt;
system_profiler SPHardwareDataType &lt;span class="o"&gt;|&lt;/span&gt; grep -E &lt;span class="s2"&gt;"Chip|Memory"&lt;/span&gt;

&lt;span class="c"&gt;# Exact byte count; divide by 1073741824 to get GB&lt;/span&gt;
sysctl -n hw.memsize
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You will get lines like &lt;code&gt;Chip: Apple ...&lt;/code&gt; and &lt;code&gt;Memory: 16 GB&lt;/code&gt;, plus a byte count such as &lt;code&gt;17179869184&lt;/code&gt;, which is 16 GB. These read the hardware directly, so you are comparing real numbers instead of marketing copy.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Run your one test task with a stopwatch.&lt;/strong&gt; Open the 20-page PDF, fire the summarise prompt, note the time, then note whether fan noise or responsiveness changes ten minutes in. One number — "finished in 40 seconds, and a second run after ten minutes was about the same" — is worth more than any chart. A short burst tells you about chat and code completion; a repeated run tells you about sustained work.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;6. Compare final prices, not starting prices.&lt;/strong&gt; Put the two totals side by side. Memory and storage upgrades are exactly where the gap between an Air-class and a Pro-class purchase opens up.&lt;/p&gt;

&lt;h2&gt;
  
  
  Air vs Pro vs Windows-only tools: a table you fill in
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Machine&lt;/th&gt;
&lt;th&gt;What to verify yourself&lt;/th&gt;
&lt;th&gt;Price&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;MacBook Air M5&lt;/td&gt;
&lt;td&gt;Not described in the source posts. Check thermal design and sustained performance on the vendor's own page&lt;/td&gt;
&lt;td&gt;Not stated in the source posts&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;MacBook Pro (M-series)&lt;/td&gt;
&lt;td&gt;Not described in the source posts. Check whether it sustains performance longer and offers larger memory options&lt;/td&gt;
&lt;td&gt;Not stated in the source posts&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Windows-only AI/GPU tool (source example: VRization)&lt;/td&gt;
&lt;td&gt;Windows + Android/iOS with GPU capture; macOS not listed — look for a macOS build first&lt;/td&gt;
&lt;td&gt;Not stated in the source posts&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Read that as a to-do list, not a verdict. Every "not stated" cell is a task for you: go to the vendor's page and fill it in yourself.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which workloads to test first
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Chat and coding assistants.&lt;/strong&gt; Burst workloads — short request, short answer. This is where an Air-class machine is usually comfortable.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local model runners.&lt;/strong&gt; The model has to fit in unified memory, so check capacity before anything else.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GPU capture and streaming tools.&lt;/strong&gt; Our VRization example is a reminder that some GPU-heavy tools never target macOS at all.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Long batch jobs.&lt;/strong&gt; Repeating the same task for an hour is where thermal design shows up — and the case worth testing in step 5.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Quick answers
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Is the MacBook Air M5 good enough for AI work in 2026?&lt;/strong&gt;&lt;br&gt;
The source posts carry no specifications, so no honest yes or no is possible from them. Decide from your own test: if every tool on your list runs on macOS and your task finishes in an acceptable time, it is good enough for that task.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do I need a MacBook Pro instead?&lt;/strong&gt;&lt;br&gt;
Only if your workload is sustained rather than bursty, or if you need more memory than the Air-class configuration offers. Verify both on the vendor's page, then compare exact-configuration prices.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can I run Windows-only AI tools like VRization on a Mac?&lt;/strong&gt;&lt;br&gt;
The source lists VRization for Windows, Android and iOS, with no macOS mention. Check for a macOS build before you spend anything.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Are there benchmarks showing the M5 beating older Macs?&lt;/strong&gt;&lt;br&gt;
Not in our source material — there are no performance figures there at all. Use the two Terminal commands and a stopwatch instead of trusting a chart.&lt;/p&gt;

&lt;p&gt;This article was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt; — and the channel is worth a follow if you want the source digests behind pieces like this one: &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;The label on the lid decides almost nothing. Two questions settle it: does your stack run on macOS, and do your models fit in the memory you are paying for. Borrow a machine for an hour, run one real task end to end, repeat it after ten minutes, and compare the final price of the exact configuration you tested. That hour costs far less than a return shipment.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>apple</category>
      <category>productivity</category>
      <category>tutorial</category>
    </item>
    <item>
      <title>Enabling Hidden Windows 11 Features with ViVeTool Without Breaking Anything</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Thu, 12 Nov 2026 10:09:33 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/enabling-hidden-windows-11-features-with-vivetool-without-breaking-anything-2nbk</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/enabling-hidden-windows-11-features-with-vivetool-without-breaking-anything-2nbk</guid>
      <description>&lt;p&gt;Windows 11 ships with code for features that are compiled but switched off. ViVeTool is the utility that flips those switches, and most guides make it look like a one-liner. In practice the &lt;code&gt;/enable&lt;/code&gt; command is the easy part — the work is in the three checks around it: does your build actually contain the feature, does the ID match your build, and can you get back to where you started?&lt;/p&gt;

&lt;p&gt;This is the workflow I use. Budget 20–30 minutes, including a reboot and a restore-point check. It's an intermediate task; if you've never opened an elevated PowerShell window, read through once before running anything.&lt;/p&gt;

&lt;h2&gt;
  
  
  A feature flag is a switch, not a delivery truck
&lt;/h2&gt;

&lt;p&gt;A flag is a hidden on/off bit inside Windows. ViVeTool changes the bit. It cannot add code that isn't there.&lt;/p&gt;

&lt;p&gt;That single sentence explains most of the frustration people hit. If the feature code hasn't landed in your build, &lt;code&gt;/enable&lt;/code&gt; may still print a success message while nothing observable changes. Feature IDs are build-specific, so an ID copied from a blog post about a different build is not a shortcut — it's a dead end. Treat every ID as unverified until &lt;code&gt;/query&lt;/code&gt; confirms it exists on &lt;em&gt;your&lt;/em&gt; machine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prep work that actually matters
&lt;/h2&gt;

&lt;p&gt;Gather these before you start:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Windows 11 with a known build number (you'll record it in the next step).&lt;/li&gt;
&lt;li&gt;Administrator rights — ViVeTool writes system configuration and normally wants an elevated prompt.&lt;/li&gt;
&lt;li&gt;A ViVeTool release matching your CPU architecture, extracted to a local folder, e.g. &lt;code&gt;C:\Users\&amp;lt;you&amp;gt;\Downloads\ViVeTool\ViVeTool.exe&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;The target feature ID, taken from the release notes or a trusted tracker for your exact build. I'll use &lt;code&gt;&amp;lt;feature-id&amp;gt;&lt;/code&gt; as a placeholder throughout.&lt;/li&gt;
&lt;li&gt;System Protection enabled, or permission to create a restore point.&lt;/li&gt;
&lt;li&gt;Free space for the archive, the extracted files, and the checkpoint. &lt;code&gt;Get-PSDrive C&lt;/code&gt; tells you.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Create the checkpoint first:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Enable&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;ComputerRestore&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Drive&lt;/span&gt; &lt;span class="s2"&gt;"C:\"
Checkpoint-Computer -Description "&lt;/span&gt;&lt;span class="nc"&gt;Before&lt;/span&gt; &lt;span class="nc"&gt;ViVeTool&lt;/span&gt;&lt;span class="s2"&gt;" -RestorePointType MODIFY_SETTINGS
Get-ComputerRestorePoint | Select-Object -Last 1
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The last command should come back with a recent entry reading &lt;code&gt;Description : Before ViVeTool&lt;/code&gt;. If System Protection is disabled by policy, the checkpoint command fails — stop there and fix that first. Running ViVeTool without a rollback point is the one mistake that turns a 20-minute experiment into an afternoon.&lt;/p&gt;

&lt;p&gt;If the feature depends on a Windows update you haven't installed yet, install the update before doing anything else.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 1: Record the build, architecture, and rollback baseline
&lt;/h2&gt;

&lt;p&gt;Open an administrator PowerShell window and save the OS build, architecture, and your feature-ID worksheet.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nv"&gt;$baseline&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Desktop\vive-baseline.txt"&lt;/span&gt;
&lt;span class="nc"&gt;Get&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;ComputerInfo&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Select&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="nc"&gt;WindowsProductName&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;WindowsVersion&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;OsBuildNumber&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;OsArchitecture&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Tee&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;FilePath&lt;/span&gt; &lt;span class="nv"&gt;$baseline&lt;/span&gt;
&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;System&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Runtime&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;InteropServices&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;RuntimeInformation&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="o"&gt;::&lt;/span&gt;&lt;span class="nc"&gt;OSArchitecture&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Tee&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;FilePath&lt;/span&gt; &lt;span class="nv"&gt;$baseline&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Append&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Expected shape of the output:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;WindowsProductName : Windows 11 ...
WindowsVersion     : &amp;lt;version&amp;gt;
OsBuildNumber      : &amp;lt;build-number&amp;gt;
OsArchitecture     : &amp;lt;architecture&amp;gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The architecture line decides which ViVeTool build you download: &lt;code&gt;X64&lt;/code&gt; means x64, &lt;code&gt;Arm64&lt;/code&gt; means ARM64. Don't mix them.&lt;/p&gt;

&lt;p&gt;Then write a small worksheet so the experiment stays reproducible:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Item&lt;/th&gt;
&lt;th&gt;Example value&lt;/th&gt;
&lt;th&gt;Why it matters&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Feature ID&lt;/td&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;feature-id&amp;gt;&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Must match the build&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Build before&lt;/td&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;build-number&amp;gt;&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Rollback reference&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;ViVeTool SHA256&lt;/td&gt;
&lt;td&gt;&lt;code&gt;&amp;lt;sha256&amp;gt;&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Integrity check&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Result after reboot&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;Enabled&lt;/code&gt; or &lt;code&gt;No change&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;Keep-or-rollback decision&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Before continuing, confirm: build saved to &lt;code&gt;vive-baseline.txt&lt;/code&gt;, architecture checked, feature ID recorded &lt;em&gt;with its source&lt;/em&gt;, restore point visible in &lt;code&gt;Get-ComputerRestorePoint&lt;/code&gt;, and only one ID in play.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: Stage ViVeTool and enable exactly one flag
&lt;/h2&gt;

&lt;p&gt;Extract, hash, query, enable — in that order.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nv"&gt;$work&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Downloads\ViVeTool"&lt;/span&gt;
&lt;span class="nc"&gt;New&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Item&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;ItemType&lt;/span&gt; &lt;span class="nc"&gt;Directory&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Path&lt;/span&gt; &lt;span class="nv"&gt;$work&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Force&lt;/span&gt;
&lt;span class="nc"&gt;Expand&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Archive&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;LiteralPath&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Downloads\ViVeTool.zip"&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;DestinationPath&lt;/span&gt; &lt;span class="nv"&gt;$work&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Force&lt;/span&gt;
&lt;span class="nc"&gt;Set&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Location&lt;/span&gt; &lt;span class="nv"&gt;$work&lt;/span&gt;
&lt;span class="nc"&gt;Get&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;ChildItem&lt;/span&gt;
&lt;span class="nc"&gt;Get&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;FileHash&lt;/span&gt; &lt;span class="p"&gt;.&lt;/span&gt;\&lt;span class="nc"&gt;ViVeTool&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;exe&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Algorithm&lt;/span&gt; &lt;span class="no"&gt;SHA256&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Compare the SHA256 value against the one published with your ViVeTool release. If they differ, delete the folder and download again — don't reason with it.&lt;/p&gt;

&lt;p&gt;Now elevate:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Start&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="k"&gt;Process&lt;/span&gt; &lt;span class="n"&gt;powershell&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Verb&lt;/span&gt; &lt;span class="nc"&gt;RunAs&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In the elevated window, query before you change anything:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Set&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Location&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Downloads\ViVeTool"&lt;/span&gt;
&lt;span class="p"&gt;.&lt;/span&gt;\&lt;span class="nc"&gt;ViVeTool&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;exe&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;query&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;
&lt;span class="p"&gt;.&lt;/span&gt;\&lt;span class="nc"&gt;ViVeTool&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;exe&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;query&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Desktop\vive-before.txt"&lt;/span&gt;
&lt;span class="p"&gt;.&lt;/span&gt;\&lt;span class="nc"&gt;ViVeTool&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;exe&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;enable&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If &lt;code&gt;/query /id:&amp;lt;feature-id&amp;gt;&lt;/code&gt; returns no matching feature, stop. The ID is wrong for this build, and enabling it will only produce noise. Saving the full &lt;code&gt;/query&lt;/code&gt; dump to the Desktop gives you a before/after comparison later.&lt;/p&gt;

&lt;p&gt;A successful enable prints something like &lt;code&gt;Successfully set feature configuration(s)&lt;/code&gt;. Then reboot so the flag is loaded at startup:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Restart&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Computer&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Why one ID at a time? Attribution. Enable five flags, the UI breaks, and you have no way to tell which one did it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: Verify twice — state, then behavior
&lt;/h2&gt;

&lt;p&gt;After the reboot, query the same ID again and &lt;em&gt;then&lt;/em&gt; look at the UI or app the flag is supposed to change.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Set&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Location&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Downloads\ViVeTool"&lt;/span&gt;
&lt;span class="p"&gt;.&lt;/span&gt;\&lt;span class="nc"&gt;ViVeTool&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;exe&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;query&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="nc"&gt;Select&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="kt"&gt;String&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Pattern&lt;/span&gt; &lt;span class="s2"&gt;"Enabled|State|Feature"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;ViVeTool state and visible behavior are two separate checks, and this is where most write-ups quietly skip the failure case. A flag can report enabled while the UI stays identical, because a dependency, an app version, or a server-side experiment is still missing.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Query result&lt;/th&gt;
&lt;th&gt;Visible result&lt;/th&gt;
&lt;th&gt;Action&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;State shows enabled&lt;/td&gt;
&lt;td&gt;Feature appears&lt;/td&gt;
&lt;td&gt;Keep it; record build and ID&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;State shows enabled&lt;/td&gt;
&lt;td&gt;Feature absent&lt;/td&gt;
&lt;td&gt;Check dependencies, app version, server-side gating; then disable&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;No matching feature&lt;/td&gt;
&lt;td&gt;Not applicable&lt;/td&gt;
&lt;td&gt;Wrong build or wrong ID; stop and re-check the source&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;If the feature doesn't show up and you don't need it, roll back right away instead of enabling more IDs and hoping:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="p"&gt;.&lt;/span&gt;\&lt;span class="nc"&gt;ViVeTool&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;exe&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;disable&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="n"&gt;feature&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;
&lt;span class="nc"&gt;Restart&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Computer&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;After the reboot, &lt;code&gt;/query /id:&amp;lt;feature-id&amp;gt;&lt;/code&gt; should show the feature as disabled or inactive. If it doesn't, restore the checkpoint from Step 1.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two errors you will probably hit
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;&lt;code&gt;The term 'ViVeTool.exe' is not recognized...&lt;/code&gt;&lt;/strong&gt; — PowerShell won't run executables from the current directory unless you prefix them with &lt;code&gt;.\&lt;/code&gt;, or you're simply in the wrong folder. Fix both at once:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Set&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Location&lt;/span&gt; &lt;span class="s2"&gt;"$env:USERPROFILE\Downloads\ViVeTool"&lt;/span&gt;
&lt;span class="p"&gt;.&lt;/span&gt;\&lt;span class="nc"&gt;ViVeTool&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;exe&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;query&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;&lt;code&gt;Access is denied&lt;/code&gt;&lt;/strong&gt; — the window isn't elevated, so ViVeTool can't write the feature configuration. Confirm elevation before retrying:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;Security&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Principal&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;WindowsPrincipal&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;Security&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Principal&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;WindowsIdentity&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="o"&gt;::&lt;/span&gt;&lt;span class="nc"&gt;GetCurrent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;IsInRole&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;Security&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Principal&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;WindowsBuiltInRole&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="o"&gt;::&lt;/span&gt;&lt;span class="nc"&gt;Administrator&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;True&lt;/code&gt; means you're good; &lt;code&gt;False&lt;/code&gt; means start a new administrator window.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I take away
&lt;/h2&gt;

&lt;p&gt;ViVeTool is a switch-flipper, not a feature installer. The ID has to exist in your build, the flag state has to be checked separately from the UI, and every experiment needs a checkpoint behind it plus a one-command way out. Record the build, change one ID, reboot, verify twice, and keep the result only if both checks agree.&lt;/p&gt;

&lt;p&gt;This article was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt;. More practical Windows and tooling guides land regularly on the channel: &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>windows</category>
      <category>tutorial</category>
      <category>productivity</category>
      <category>security</category>
    </item>
    <item>
      <title>Labeling AI content for Russia's March 2027 rules: an engineering checklist</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Wed, 11 Nov 2026 10:12:40 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/labeling-ai-content-for-russias-march-2027-rules-an-engineering-checklist-3pln</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/labeling-ai-content-for-russias-march-2027-rules-an-engineering-checklist-3pln</guid>
      <description>&lt;p&gt;Russian rules on labeling AI-generated content take effect on 1 March 2027. The legal analysis belongs to someone else; what landed on my desk is the engineering side — making sure every asset that leaves the pipeline carries a provenance record both a person and a machine can read, and that it cannot quietly lose that record later. Below is the setup I would hand to another developer.&lt;/p&gt;

&lt;p&gt;Budget 90–120 minutes for the first asset. Once the manifest format, the metadata commands and the CI script exist, each additional asset costs roughly ten minutes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Layout and prerequisites
&lt;/h2&gt;

&lt;p&gt;Rule number one: every artifact gets one manifest and one asset file sharing the same ID. That single convention is what makes everything else scriptable — checksums, metadata writing and the CI loop all key off the same name.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;mkdir -p ai-labeling/&lt;span class="o"&gt;{&lt;/span&gt;schema,manifests,assets,scripts&lt;span class="o"&gt;}&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; ai-labeling
git init
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;ai-labeling/
├── schema/label.schema.json
├── manifests/img-2027-0001.json
├── assets/img-2027-0001.jpg
└── scripts/verify_labels.sh
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Tools to have ready, with the version check I run first:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Python 3&lt;/strong&gt; with the &lt;code&gt;jsonschema&lt;/code&gt; package (&lt;code&gt;pip install jsonschema&lt;/code&gt;); confirm with &lt;code&gt;python3 --version&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Node.js LTS&lt;/strong&gt; — &lt;code&gt;ajv-cli&lt;/code&gt; runs through &lt;code&gt;npx&lt;/code&gt;, so there is nothing to install globally: check &lt;code&gt;node -v&lt;/code&gt; and &lt;code&gt;npx --version&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;ExifTool&lt;/strong&gt; — writes metadata into image files: &lt;code&gt;exiftool -ver&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;FFmpeg&lt;/strong&gt; — writes metadata into video containers: &lt;code&gt;ffmpeg -version | head -n 1&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Linux, macOS, or Windows 10+ with WSL2. On PowerShell 7, record the shell version with &lt;code&gt;$PSVersionTable.PSVersion&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Write access to the labeling repo plus permission to run its CI pipeline.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Step 1: Write the manifest first, validate it against a schema
&lt;/h2&gt;

&lt;p&gt;The manifest is a plain-text provenance record: what produced the asset, when, and whether a human reviewed it. It exists before any output surface, so the file metadata, the page label and the API header are all derived from one source and cannot drift apart. Writing it first also forces you to capture the model identifier while you still have it — copy the exact string from the API response body, never type a model name from memory.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;manifests/img-2027-0001.json&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"asset_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"img-2027-0001"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"asset_sha256"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&amp;lt;output-of-sha256sum&amp;gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"created_at"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"2027-02-14T10:22:00Z"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"origin"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"ai-generated"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"model_provider"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&amp;lt;provider-name-from-your-contract&amp;gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"model_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&amp;lt;exact-model-id-returned-by-the-api&amp;gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"ai_generated_share"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mf"&gt;1.0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"human_reviewed"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"reviewer"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"editor@example.com"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"reviewed_at"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"2027-02-14T11:05:00Z"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"disclosure"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"visible_label"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"AI-generated image"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"machine_readable"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For AI-written text that an editor reworked, keep the same schema and set &lt;code&gt;origin&lt;/code&gt; to &lt;code&gt;ai-assisted&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Describe the fields in &lt;code&gt;schema/label.schema.json&lt;/code&gt; and make these required: &lt;code&gt;asset_id&lt;/code&gt;, &lt;code&gt;asset_sha256&lt;/code&gt;, &lt;code&gt;created_at&lt;/code&gt;, &lt;code&gt;origin&lt;/code&gt;, &lt;code&gt;model_id&lt;/code&gt;, &lt;code&gt;ai_generated_share&lt;/code&gt;, &lt;code&gt;human_reviewed&lt;/code&gt;, &lt;code&gt;disclosure&lt;/code&gt;. Then hash the file and validate:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;sha256sum assets/img-2027-0001.jpg &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; checksums.txt
python3 -c &lt;span class="s2"&gt;"import json,jsonschema;jsonschema.validate(json.load(open('manifests/img-2027-0001.json')),json.load(open('schema/label.schema.json')));print('manifest valid')"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Confirm the schema actually enforces something:&lt;/strong&gt; the terminal should print &lt;code&gt;manifest valid&lt;/code&gt;. Then copy the manifest, remove &lt;code&gt;model_id&lt;/code&gt;, validate the copy, and check that you get a &lt;code&gt;ValidationError&lt;/code&gt; naming that field. A schema that never fails is decoration.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: Three surfaces, not one
&lt;/h2&gt;

&lt;p&gt;Three different readers need three different surfaces. Metadata inside the file survives download and re-upload. The visible label is what a human sees. An HTTP header is what an automated client sees. A caption alone vanishes the moment someone embeds the image elsewhere, and metadata alone is invisible to a reader — ship both.&lt;/p&gt;

&lt;p&gt;Image metadata with ExifTool (&lt;code&gt;-overwrite_original&lt;/code&gt; keeps &lt;code&gt;.jpg_original&lt;/code&gt; backups out of your build output):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;exiftool -overwrite_original \
  -XMP-dc:Description=&lt;span class="s2"&gt;"AI-generated image. Model: &amp;lt;model-id&amp;gt;. Provenance: /provenance/img-2027-0001.json"&lt;/span&gt; \
  -XMP-xmpMM:DocumentID=&lt;span class="s2"&gt;"img-2027-0001"&lt;/span&gt; \
  assets/img-2027-0001.jpg
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For video, &lt;code&gt;-codec copy&lt;/code&gt; remuxes without re-encoding, so quality is untouched and the job finishes in seconds:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;ffmpeg -i assets/clip-2027-0002.mp4 -codec copy \
  -metadata &lt;span class="nv"&gt;comment&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"AI-generated video; provenance clip-2027-0002"&lt;/span&gt; \
  assets/clip-2027-0002.labeled.mp4
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Page markup — one tag for machines, one sentence for people:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight html"&gt;&lt;code&gt;&lt;span class="nt"&gt;&amp;lt;meta&lt;/span&gt; &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s"&gt;"ai-generated"&lt;/span&gt; &lt;span class="na"&gt;content&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s"&gt;"true"&lt;/span&gt;&lt;span class="nt"&gt;&amp;gt;&lt;/span&gt;
&lt;span class="nt"&gt;&amp;lt;p&lt;/span&gt; &lt;span class="na"&gt;class&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s"&gt;"ai-disclosure"&lt;/span&gt;&lt;span class="nt"&gt;&amp;gt;&lt;/span&gt;AI-generated image. &lt;span class="nt"&gt;&amp;lt;a&lt;/span&gt; &lt;span class="na"&gt;href&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s"&gt;"/provenance/img-2027-0001.json"&lt;/span&gt;&lt;span class="nt"&gt;&amp;gt;&lt;/span&gt;Provenance record&lt;span class="nt"&gt;&amp;lt;/a&amp;gt;&amp;lt;/p&amp;gt;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And on every API endpoint that returns generated text, image or audio:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;X-AI-Generated: true
X-AI-Provenance: /provenance/img-2027-0001.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If your provider supports embedded, signed provenance manifests, sign the file with &lt;code&gt;c2patool&lt;/code&gt; instead of plain XMP. Flag names and signing-key setup change between releases, so read the current docs and pin the version you used.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Check both surfaces:&lt;/strong&gt; &lt;code&gt;exiftool -s3 -XMP-dc:Description assets/img-2027-0001.jpg&lt;/code&gt; should print the description line, and &lt;code&gt;curl -sI https://&amp;lt;your-domain&amp;gt;/api/render | grep -i x-ai-generated&lt;/code&gt; should return &lt;code&gt;x-ai-generated: true&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: One verification script, required in CI
&lt;/h2&gt;

&lt;p&gt;Labels disappear quietly. Design tools, image optimizers and CDN pipelines strip XMP blocks without complaint, and nobody notices until someone else does. A failing build is the only reminder that survives a team change.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;scripts/verify_labels.sh&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;#!/usr/bin/env bash&lt;/span&gt;
&lt;span class="nb"&gt;set&lt;/span&gt; -euo pipefail
npx ajv-cli validate -s schema/label.schema.json -d &lt;span class="s2"&gt;"manifests/*.json"&lt;/span&gt; --errors=text
sha256sum -c checksums.txt
&lt;span class="k"&gt;for&lt;/span&gt; f &lt;span class="k"&gt;in&lt;/span&gt; assets/*.jpg&lt;span class="o"&gt;;&lt;/span&gt; &lt;span class="k"&gt;do&lt;/span&gt;
  exiftool -s3 -XMP-dc:Description &lt;span class="s2"&gt;"$f"&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; grep -q &lt;span class="s2"&gt;"AI-generated"&lt;/span&gt; \
    &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="o"&gt;{&lt;/span&gt; &lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s2"&gt;"MISSING LABEL: $f"&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt; &lt;span class="nb"&gt;exit&lt;/span&gt; 1&lt;span class="o"&gt;;&lt;/span&gt; &lt;span class="o"&gt;}&lt;/span&gt;
&lt;span class="k"&gt;done&lt;/span&gt;
&lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s2"&gt;"all labels verified"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Wire it into CI as a single step running &lt;code&gt;bash scripts/verify_labels.sh&lt;/code&gt;, and make that step required for merges.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Check it end to end:&lt;/strong&gt; three lines should appear in order — &lt;code&gt;manifests/img-2027-0001.json valid&lt;/code&gt;, &lt;code&gt;assets/img-2027-0001.jpg: OK&lt;/code&gt;, &lt;code&gt;all labels verified&lt;/code&gt;. Then copy an asset, strip it with &lt;code&gt;exiftool -all= assets/test-unlabeled.jpg&lt;/code&gt;, re-run the script and confirm you see &lt;code&gt;MISSING LABEL: assets/test-unlabeled.jpg&lt;/code&gt; followed by &lt;code&gt;echo $?&lt;/code&gt; printing &lt;code&gt;1&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the script cannot decide
&lt;/h2&gt;

&lt;p&gt;Scope is a legal question, not a build question. Which of your outputs fall under the rules, which exemptions apply, and the exact wording of the disclosure all have to be confirmed against the current regulation text with your legal team before you ship anything. Everything above covers only the mechanics of producing and enforcing the label.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;One manifest per asset, same ID as the asset, validated against a schema.&lt;/li&gt;
&lt;li&gt;Copy &lt;code&gt;model_id&lt;/code&gt; from the API response at generation time, not from memory.&lt;/li&gt;
&lt;li&gt;Write the label into the file, the page and the API response.&lt;/li&gt;
&lt;li&gt;Verify with a single script, run it in CI, make it a required check.&lt;/li&gt;
&lt;li&gt;Pin your tool versions; c2patool syntax and key setup move between releases.&lt;/li&gt;
&lt;li&gt;Get the legal scope sign-off in writing before the first deploy.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This article was originally published on AI Release — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt;. Shorter, recurring notes on tooling and compliance work like this show up in the channel at &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>tutorial</category>
      <category>programming</category>
      <category>devops</category>
    </item>
    <item>
      <title>MCP in practice: connecting your first server to an AI agent</title>
      <dc:creator>AI Release</dc:creator>
      <pubDate>Wed, 11 Nov 2026 10:05:09 +0000</pubDate>
      <link>https://dev.to/qunirice_verase_1ab0caeba/mcp-in-practice-connecting-your-first-server-to-an-ai-agent-565i</link>
      <guid>https://dev.to/qunirice_verase_1ab0caeba/mcp-in-practice-connecting-your-first-server-to-an-ai-agent-565i</guid>
      <description>&lt;p&gt;Most conversations I have about MCP start with the same misunderstanding: people expect the model itself to &lt;em&gt;do&lt;/em&gt; something. It can't, not on its own. The Model Context Protocol is the layer that closes that gap — it lets an AI assistant reach something outside the chat window, whether that's an Android phone or an app store account. Below is how the pieces fit together and how I'd wire up a first server without getting lost.&lt;/p&gt;

&lt;h2&gt;
  
  
  A brain and a pair of hands
&lt;/h2&gt;

&lt;p&gt;The Pony project puts the idea better than I can: Pony is the hands, you pick the brain. Your model decides &lt;em&gt;what&lt;/em&gt; should happen; the MCP server is what actually executes it on the device or account. Once I started thinking in those two halves, most of the setup confusion disappeared. The model you pick sets how well the agent reasons. The server you pick sets what it is allowed to touch.&lt;/p&gt;

&lt;p&gt;That split also explains why "which model supports MCP?" is the wrong first question. Claude, Grok, Gemini and any other MCP-capable agent can drive the same server — you are choosing a brain, not a platform lock-in.&lt;/p&gt;

&lt;h2&gt;
  
  
  The checklist before you touch anything
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;An AI brain: Claude, Grok, Gemini, or any MCP agent&lt;/li&gt;
&lt;li&gt;Hands: a server that performs real actions&lt;/li&gt;
&lt;li&gt;The target ready — an Android phone, App Store Connect, Google Play Console&lt;/li&gt;
&lt;li&gt;An MCP client that can launch and connect to servers&lt;/li&gt;
&lt;li&gt;Credentials for whatever external service you're connecting&lt;/li&gt;
&lt;li&gt;A clear-eyed acceptance of the trade-off: you are handing an agent direct access&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That last point is not a formality. Once permissions are granted, the server does what the model asks. Decide up front what you're comfortable with.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three servers worth knowing
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Pony&lt;/strong&gt; — hands an Android phone to an AI agent as a remote control. It needs a physical device.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;shipstores&lt;/strong&gt; — bridges App Store Connect, Google Play Console and Expo EAS, so an agent can build and publish iOS and Android apps.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;magaza-mcp&lt;/strong&gt; — a single server covering both stores, with the entire interface in Turkish.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Pick the one that matches your problem before you start installing anything.&lt;/p&gt;

&lt;h2&gt;
  
  
  Wiring it up, step by step
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Choose the brain.&lt;/strong&gt; Start with a model that speaks MCP. This decides the quality of the decisions, not the reach.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Choose the hands.&lt;/strong&gt; Pony for a device, shipstores or magaza-mcp for store publishing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Install the server on the target side.&lt;/strong&gt; For Pony that means running it on the Android phone itself; for the store servers, a machine that can reach your store accounts. You're done when the server is up and waiting for a connection.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Register it in your client.&lt;/strong&gt; Open the configuration area of your AI client and add the server address or startup command. You're done when the client discovers the server and lists its available tools.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Grant permissions.&lt;/strong&gt; Pony needs accessibility or comparable device-level access. The store servers need your App Store Connect and Google Play Console accounts connected, plus the API keys those accounts require.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Run a small test.&lt;/strong&gt; Keep it boring: "open the camera" or "check the latest app build status." If a real answer comes back from the device or the store, the connection works.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What usually goes wrong
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;The agent can't see the server.&lt;/strong&gt; In my experience this is almost always ordering. Start the server first, then open the MCP client — a server that comes up afterwards often goes undetected. Restart both and double-check the path in your config.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Nothing happens, silently.&lt;/strong&gt; On Android that usually means missing device-level permissions. With the store servers it means the accounts aren't properly connected or a key is absent. A silent failure is nearly always a credential failure.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The server answers, but with store errors.&lt;/strong&gt; For shipstores, confirm your account's access level actually allows publishing. For magaza-mcp, remember the interface is Turkish throughout — a language mismatch can mask an account-local problem.&lt;/p&gt;

&lt;h2&gt;
  
  
  Quick answers
&lt;/h2&gt;

&lt;p&gt;&lt;em&gt;Is one server enough for both stores?&lt;/em&gt; Yes — shipstores covers App Store Connect and Google Play Console together, and magaza-mcp does the same in a single server.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Do I need a real phone for Pony?&lt;/em&gt; Yes. Giving an agent an Android smartphone over MCP is the entire point of the project.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Does the client matter?&lt;/em&gt; It matters as a launcher: without a client that can start and connect to MCP servers, the server just sits there.&lt;/p&gt;

&lt;p&gt;If you want to dig further into agents that use this protocol in day-to-day work, there is a longer write-up on AI Release covering automation agents — &lt;a href="https://ai-release.org" rel="noopener noreferrer"&gt;https://ai-release.org&lt;/a&gt; — and new guides land on the channel at &lt;a href="https://t.me/ai_release1" rel="noopener noreferrer"&gt;https://t.me/ai_release1&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaway
&lt;/h2&gt;

&lt;p&gt;MCP is not magic and it is not complicated: it is a contract between a brain that decides and a server that acts. Choose both deliberately, install the server before you open the client, grant only the access you actually need, and validate with one trivial instruction. If "open the camera" returns a real result, you have connected your first MCP server — and everything after that is just adding more hands.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>tutorial</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
