<?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: Anton “Admiral” Fishman</title>
    <description>The latest articles on DEV Community by Anton “Admiral” Fishman (@anton_admiralfishman_1).</description>
    <link>https://dev.to/anton_admiralfishman_1</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%2F4248568%2F8a746d58-1f1e-4169-ad71-d81a79e06b4a.jpg</url>
      <title>DEV Community: Anton “Admiral” Fishman</title>
      <link>https://dev.to/anton_admiralfishman_1</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/anton_admiralfishman_1"/>
    <language>en</language>
    <item>
      <title>Local LLM generation on a consumer GPU: AirColibri streams Qwen3.5-9B into modest VRAM</title>
      <dc:creator>Anton “Admiral” Fishman</dc:creator>
      <pubDate>Tue, 17 Nov 2026 22:28:52 +0000</pubDate>
      <link>https://dev.to/anton_admiralfishman_1/local-llm-generation-on-a-consumer-gpu-aircolibri-streams-qwen35-9b-into-modest-vram-50be</link>
      <guid>https://dev.to/anton_admiralfishman_1/local-llm-generation-on-a-consumer-gpu-aircolibri-streams-qwen35-9b-into-modest-vram-50be</guid>
      <description>&lt;p&gt;I run language models at home on a mid-range NVIDIA card. A 9B-class hybrid model still asks for more memory than a lot of cards have if you load full-precision (or even comfortably quantized) weights all at once. Off-the-shelf servers are excellent when the model fits. I wanted a path that still works when it does not: measure free VRAM, keep as many layers resident as that budget allows, and stream the rest.&lt;/p&gt;

&lt;p&gt;The result is a stack for local LLM generation on a consumer GPU. &lt;strong&gt;&lt;a href="https://github.com/DIIIUA/air-colibri" rel="noopener noreferrer"&gt;AirColibri&lt;/a&gt;&lt;/strong&gt; is a Python inference engine for a single NVIDIA card. I wanted to run a hybrid &lt;strong&gt;Qwen3.5&lt;/strong&gt; model locally when the card does not have enough VRAM to keep every layer resident, and to serve it on the LAN through an OpenAI-compatible API with a small web chat.&lt;/p&gt;

&lt;p&gt;This article is about how it is put together, what I traded away, and how to try it. The repository is a new public release (mid-November 2026). AirColibri is &lt;strong&gt;1.0.0&lt;/strong&gt;; the project NOTICE still describes it as an experimental integration, and I treat the quality and speed numbers below as measurements from my own machine, not as a promise about yours.&lt;/p&gt;

&lt;h2&gt;
  
  
  Layer streaming instead of a bigger card
&lt;/h2&gt;

&lt;p&gt;AirColibri is aimed at &lt;strong&gt;Qwen3.5&lt;/strong&gt; hybrid models (Gated DeltaNet + attention) on one NVIDIA GPU, with an OpenAI-compatible server for the local network. It combines ideas from &lt;a href="https://github.com/lyogavin/airllm" rel="noopener noreferrer"&gt;AirLLM&lt;/a&gt; (layer streaming, selective weight reads) and &lt;a href="https://github.com/JustVugg/colibri" rel="noopener noreferrer"&gt;Colibrì&lt;/a&gt; (tiered placement, native MTP speculation) with its own W4A16 weights (Q4, group size 64) and NVRTC CUDA kernels. It is an independent project under &lt;strong&gt;Apache-2.0&lt;/strong&gt;; those names do not imply endorsement.&lt;/p&gt;

&lt;p&gt;The design choice I care about most is that &lt;strong&gt;standard Hugging Face shards are read directly&lt;/strong&gt; (&lt;code&gt;model-0000X-of-0000N.safetensors&lt;/code&gt; plus &lt;code&gt;model.safetensors.index.json&lt;/code&gt;). There is no converter and no second full-size copy of the model on disk. On first start, Q4 weights are built once into &lt;code&gt;data/q4-g64-cache/&lt;/code&gt;. The recommended model is &lt;a href="https://huggingface.co/Qwen/Qwen3.5-9B" rel="noopener noreferrer"&gt;&lt;code&gt;Qwen/Qwen3.5-9B&lt;/code&gt;&lt;/a&gt;; the first download is about 19 GB. You can also point &lt;code&gt;--model&lt;/code&gt; at a local folder, or drop a supported folder into &lt;code&gt;models/&lt;/code&gt;. Unsupported layouts (the README mentions OLMoE) are skipped. Colibrì per-layer containers still work.&lt;/p&gt;

&lt;p&gt;VRAM planning is measured, not profiled. At start the engine looks at &lt;strong&gt;free&lt;/strong&gt; VRAM. &lt;code&gt;--vram-mode normal&lt;/code&gt; (the default) uses almost all of it, but keeps a reserve for the OS, the KV / DeltaNet state of &lt;code&gt;--max-context&lt;/code&gt; tokens, and the prefill workspace, so a long prompt is less likely to OOM. &lt;code&gt;--vram-mode turbo&lt;/code&gt; keeps only a minimal reserve. If the whole model fits, every layer stays resident. If it does not, the remaining budget is split into equal slots that rotate roles — computing, cleaning, receiving the next layer's weights — so the next layer is already on the GPU when the current one finishes, and as many layers as fit stay resident. Prompts are prefilled in chunks (&lt;code&gt;--prefill-chunk&lt;/code&gt;, default 512 tokens), and only the last position is projected onto the 248k-token vocabulary.&lt;/p&gt;

&lt;p&gt;The LAN server binds web chat and the API on port &lt;strong&gt;8787&lt;/strong&gt;, with a UDP beacon on &lt;strong&gt;8788&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;POST /v1/chat/completions&lt;/code&gt; — OpenAI-compatible, JSON or &lt;code&gt;stream: true&lt;/code&gt; SSE (&lt;code&gt;reasoning_content&lt;/code&gt; for &lt;code&gt;&amp;lt;think&amp;gt;&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;POST /v1/chat/stop&lt;/code&gt; — stops the running generation immediately&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;GET /v1/models&lt;/code&gt;, &lt;code&gt;GET /api/info&lt;/code&gt; — model list and node status&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;GET /&lt;/code&gt; — the built-in web chat&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Only one generation runs at a time. A new request, or a client that drops its connection (including during prefill), cancels the current one and frees VRAM before the next starts. &lt;code&gt;--no-preempt&lt;/code&gt; queues instead. Bodies larger than &lt;code&gt;--max-request-bytes&lt;/code&gt; (default 8 MiB) get HTTP 413. The server is open on the LAN by default; &lt;code&gt;--require-auth&lt;/code&gt; turns on a Bearer API key. Logs go to &lt;code&gt;logs/server.log&lt;/code&gt; (rotating, 10 MB × 5).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Tradeoffs, including numbers I actually measured.&lt;/strong&gt; Q4 is lossy. Against the original weights, Q4 (W4A16, group 64) logits had a cosine similarity of about &lt;strong&gt;0.987–0.993&lt;/strong&gt; in my checks. "Bit-exact" applies only to the fused kernels versus the reference Q4 computation, not versus BF16. MTP speculation does not change greedy output: with greedy decoding the text is identical with and without MTP (&lt;code&gt;tests/test_engine_gpu.py&lt;/code&gt; covers this).&lt;/p&gt;

&lt;p&gt;On an &lt;strong&gt;RTX 3060 Ti 8 GB&lt;/strong&gt;, Qwen3.5-9B, all layers resident, 400-token answers, &lt;strong&gt;one answer at a time&lt;/strong&gt;, I measured about &lt;strong&gt;31 tok/s&lt;/strong&gt; without MTP, and &lt;strong&gt;30–37 tok/s&lt;/strong&gt; with MTP, depending on how many drafts were accepted (43–71 % in those runs; on new text MTP may give no gain). That is not a throughput number for several parallel requests, and I am not quoting one. &lt;code&gt;--fast-vocab&lt;/code&gt; is experimental and off by default; it can drop rare tokens on long outputs.&lt;/p&gt;

&lt;p&gt;The runtime I document is narrower than "any Python box": NVIDIA GPU with CUDA, Python 3.12, PyTorch 2.6 with CUDA 12.4, Windows x64 or Linux. I tested on the 3060 Ti 8 GB; the README states that cards with around 4 GB stream part of the layers.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Install and run.&lt;/strong&gt; There is no PyPI package. On Windows, &lt;code&gt;aircolibri.cmd&lt;/code&gt; picks &lt;code&gt;.venv&lt;/code&gt; automatically; on any OS you can run &lt;code&gt;python aircolibri.py&lt;/code&gt; with the same subcommands. The README's virtualenv paths are Windows-style; on Linux, use &lt;code&gt;.venv/bin/python&lt;/code&gt; in place of &lt;code&gt;.venv/Scripts/python&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;git clone https://github.com/DIIIUA/air-colibri.git
&lt;span class="nb"&gt;cd&lt;/span&gt; air-colibri
python -m venv .venv
.venv/Scripts/python -m pip install &lt;span class="nv"&gt;torch&lt;/span&gt;&lt;span class="o"&gt;==&lt;/span&gt;2.6.0 --index-url https://download.pytorch.org/whl/cu124
.venv/Scripts/python -m pip install -r requirements.txt

&lt;span class="c"&gt;# Windows: aircolibri.cmd&lt;/span&gt;
&lt;span class="c"&gt;# Any OS:  python aircolibri.py …&lt;/span&gt;
aircolibri doctor --model Qwen/Qwen3.5-9B   &lt;span class="c"&gt;# downloads once (~19 GB) and checks GPU, cache, LAN addresses&lt;/span&gt;
aircolibri ask &lt;span class="s2"&gt;"Who are you?"&lt;/span&gt;
aircolibri chat                             &lt;span class="c"&gt;# Ctrl+C stops the current answer&lt;/span&gt;
aircolibri serve                            &lt;span class="c"&gt;# web chat + OpenAI API on :8787&lt;/span&gt;
aircolibri discover                         &lt;span class="c"&gt;# find servers on the LAN&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;CPU tests: &lt;code&gt;python -m pytest tests -q&lt;/code&gt;. The GPU engine tests and the model-dependent Q4 tests run only when a CUDA GPU and a local model are present.&lt;/p&gt;

&lt;p&gt;If you already use an OpenAI-compatible client, point it at &lt;code&gt;http://&amp;lt;lan-ip&amp;gt;:8787&lt;/code&gt;. If the server is reachable without &lt;code&gt;--require-auth&lt;/code&gt;, you should treat the LAN as trusted.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is still rough
&lt;/h2&gt;

&lt;p&gt;AirColibri 1.0.0 is a public release of a deliberately experimental stack: Q4 quantization, layer streaming, and a single-flight server. It will not match a fully resident, less lossy runtime on a card that can hold the model comfortably. It is CUDA-only. The first model download is large. I have not published numbers for parallel request throughput, and I have not claimed results on GPUs I have not measured.&lt;/p&gt;

&lt;p&gt;I also will not invent a head-to-head against llama.cpp, vLLM, or Ollama. If your model already fits in one of those, use them.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would like to hear back
&lt;/h2&gt;

&lt;p&gt;Quality of Q4 Qwen3.5-9B on real tasks, not just logit cosine; behaviour on cards around 4 GB, where layers have to stream; MTP accept rates on your prompts; and whether &lt;code&gt;POST /v1/chat/completions&lt;/code&gt; and &lt;code&gt;/v1/chat/stop&lt;/code&gt; behave correctly in the clients you use.&lt;/p&gt;

&lt;p&gt;Issues and pull requests are welcome. Please do not open a PR that includes API keys, hostnames, or device identifiers.&lt;/p&gt;

&lt;h2&gt;
  
  
  Links
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;AirColibri (Python, Apache-2.0, v1.0.0): &lt;a href="https://github.com/DIIIUA/air-colibri" rel="noopener noreferrer"&gt;github.com/DIIIUA/air-colibri&lt;/a&gt; · &lt;a href="https://github.com/DIIIUA/air-colibri/releases/tag/v1.0.0" rel="noopener noreferrer"&gt;Release v1.0.0&lt;/a&gt; · recommended model &lt;a href="https://huggingface.co/Qwen/Qwen3.5-9B" rel="noopener noreferrer"&gt;Qwen/Qwen3.5-9B&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I am Anton (&lt;a href="https://github.com/DIIIUA" rel="noopener noreferrer"&gt;DIIIUA&lt;/a&gt;). If you try this on your hardware, I am more interested in a failed setup than in a compliment.&lt;/p&gt;

</description>
      <category>python</category>
      <category>llm</category>
      <category>cuda</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>An AI app for Android: ForgeMind, a Kotlin workstation for cloud and local models</title>
      <dc:creator>Anton “Admiral” Fishman</dc:creator>
      <pubDate>Tue, 17 Nov 2026 22:26:56 +0000</pubDate>
      <link>https://dev.to/anton_admiralfishman_1/an-ai-app-for-android-forgemind-a-kotlin-workstation-for-cloud-and-local-models-195a</link>
      <guid>https://dev.to/anton_admiralfishman_1/an-ai-app-for-android-forgemind-a-kotlin-workstation-for-cloud-and-local-models-195a</guid>
      <description>&lt;p&gt;I work from a phone more often than I would like to admit. Most mobile AI apps I tried were front-ends: one provider, a prompt box, maybe a document picker. As soon as I needed a shell on a VPS, a file on the phone, an MCP server at home, or the ability to tap something in another app, I was back at a computer.&lt;/p&gt;

&lt;p&gt;So I built an AI app for Android. &lt;strong&gt;&lt;a href="https://github.com/DIIIUA/ForgeMind" rel="noopener noreferrer"&gt;ForgeMind&lt;/a&gt;&lt;/strong&gt; is a native Kotlin + Jetpack Compose app. I wanted a workstation on the device, not another thin chat client: cloud and local models, SSH (including Cloudflare Access), MCP in both directions, files and a shell, and a way to read and operate the screen of other apps.&lt;/p&gt;

&lt;p&gt;This article is about how it is put together, what I traded away, and how to try it. The repository is a new public release (mid-November 2026). ForgeMind is &lt;strong&gt;0.2.1-alpha&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  A dual-contour hub on Android
&lt;/h2&gt;

&lt;p&gt;The short version of the architecture is in the README. The idea is a &lt;strong&gt;dual contour&lt;/strong&gt; — external connections and on-device ones — without shipping a multi-gigabyte Linux container inside the APK.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Outside the phone&lt;/strong&gt;, the app speaks five native LLM wire protocols: OpenAI-compatible &lt;code&gt;/chat/completions&lt;/code&gt; (SSE, with a non-SSE fallback), OpenAI &lt;code&gt;/responses&lt;/code&gt;, Anthropic &lt;code&gt;/v1/messages&lt;/code&gt;, Gemini &lt;code&gt;streamGenerateContent&lt;/code&gt;, and Ollama's native &lt;code&gt;/api/chat&lt;/code&gt;. There are twelve cloud presets (OpenAI, Anthropic, Gemini, OpenRouter, DeepSeek, xAI Grok, Groq, Mistral, Venice AI, NVIDIA NIM, Cloudflare Workers AI, Hugging Face), plus arbitrary custom HTTP gateways (six auth modes, custom headers and body, optional insecure TLS for LAN, DNS-over-HTTPS fallback), plus a LAN scanner. You can probe a single &lt;code&gt;IP:port&lt;/code&gt; in about a second, or scan the &lt;code&gt;/24&lt;/code&gt; for servers such as Ollama, LM Studio, llama.cpp, KoboldCpp, vLLM, SGLang, LocalAI, Jan, GPT4All, OpenWebUI, and LiteLLM.&lt;/p&gt;

&lt;p&gt;SSH and SFTP use JSch, with password or key auth, over a public IP, a VPN, or Tailscale. For Zero Trust I implemented Cloudflare Access SSH over &lt;code&gt;wss://&lt;/code&gt;: the app sends &lt;code&gt;CF-Access-Client-Id&lt;/code&gt; and &lt;code&gt;CF-Access-Client-Secret&lt;/code&gt; and does not need a &lt;code&gt;cloudflared&lt;/code&gt; binary on Android, and the server does not need an open SSH port.&lt;/p&gt;

&lt;p&gt;MCP is bidirectional. As a client, ForgeMind supports Streamable HTTP (&lt;code&gt;POST /mcp&lt;/code&gt;), HTTP+SSE, and &lt;code&gt;SSH_STDIO&lt;/code&gt; (it can spawn &lt;code&gt;npx&lt;/code&gt;, &lt;code&gt;uvx&lt;/code&gt;, or &lt;code&gt;python3&lt;/code&gt; MCP servers on a remote host or in Termux). As a host, a built-in Streamable HTTP server listens on &lt;code&gt;127.0.0.1:8765&lt;/code&gt; (&lt;code&gt;/mcp&lt;/code&gt; and &lt;code&gt;/health&lt;/code&gt;, Bearer token, toggles for write/shell) so a local agent, Termux, &lt;code&gt;adb forward&lt;/code&gt;, or a tunnel can call the phone's tools.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;On the device&lt;/strong&gt;, four file environments share a dual-pane manager: the app sandbox, shared phone storage, Storage Access Framework folders, and SFTP. The Sora editor shows unified diffs and takes a rollback snapshot before an edit. A persistent &lt;code&gt;/system/bin/sh&lt;/code&gt; keeps its working directory if SSH drops. Termux is a first-class contour (&lt;code&gt;127.0.0.1:2222&lt;/code&gt; and &lt;code&gt;RUN_COMMAND&lt;/code&gt;). There is an 8-window "Turbo" agent browser, separate from the ordinary in-app browser, and a floating chat overlay.&lt;/p&gt;

&lt;p&gt;The part I get asked about most is screen control. After you opt in to the Accessibility service, ForgeMind can read the visible text and controls of other apps as a tagged tree (&lt;code&gt;[#1]…&lt;/code&gt;, windows &lt;code&gt;[W1]…&lt;/code&gt;, including split screen, floating windows, and the IME), then tap, type, or scroll by tag, including a batch of actions in one call. Screenshots are on demand (Android 11+), not a background recording. Android still hides password fields, and &lt;code&gt;FLAG_SECURE&lt;/code&gt; still blocks screenshots. I call this out because "the agent can use your phone" should not sound magical.&lt;/p&gt;

&lt;p&gt;Under the chat loop there are &lt;strong&gt;56 built-in tools&lt;/strong&gt; (29 core + 27 in-app levers), plus whatever an attached MCP server adds. For small local GGUF models the schema is compacted automatically, and the parser understands seven inline tool-call syntaxes. The model picker can auto-detect a loaded context window (LM Studio, llama.cpp, Ollama, &lt;code&gt;/v1/models&lt;/code&gt;) and override it with one tap.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Tradeoffs I accepted.&lt;/strong&gt; This is an alpha. The first public tag was &lt;code&gt;v0.1.0-alpha&lt;/code&gt; on 13 November 2026; &lt;code&gt;v0.2.1-alpha&lt;/code&gt; followed a day later, mainly to stop the model picker crashing on aggregator catalogs with more than a thousand models. Accessibility is powerful and easy to get wrong on OEM skins (the README already mentions HyperOS / MIUI overlays). I encrypt secrets on the device with AES-256-GCM and do not ship API keys in the repo, but you are still putting a tool-using agent on a phone — the approval toggles and destructive-shell intercepts are there for a reason. The project is &lt;strong&gt;GPL-3.0&lt;/strong&gt;; I retain the right to ship official builds under a separate commercial license.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Requirements:&lt;/strong&gt; Android 10+ (API 29, compile SDK 36), Kotlin 2.2.21. I currently ship a prerelease APK from &lt;a href="https://github.com/DIIIUA/ForgeMind/releases/latest" rel="noopener noreferrer"&gt;GitHub Releases&lt;/a&gt; (&lt;code&gt;ForgeMind-v0.2.1-alpha.apk&lt;/code&gt;). You can also add &lt;code&gt;https://github.com/DIIIUA/ForgeMind&lt;/code&gt; in &lt;a href="https://github.com/ImranR98/Obtainium" rel="noopener noreferrer"&gt;Obtainium&lt;/a&gt; with &lt;em&gt;Include prereleases&lt;/em&gt; enabled, or build from source:&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 https://github.com/DIIIUA/ForgeMind.git
&lt;span class="nb"&gt;cd&lt;/span&gt; ForgeMind
./gradlew testDebugUnitTest assembleDebug
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That Gradle invocation is also how I run the 96 unit tests. After install: &lt;strong&gt;Settings → AI Providers&lt;/strong&gt; for a cloud key or a custom endpoint; &lt;strong&gt;Find Local AIs in Network&lt;/strong&gt; for LAN servers; &lt;strong&gt;Environments&lt;/strong&gt; for Termux or Cloudflare SSH; &lt;strong&gt;MCP Servers&lt;/strong&gt; for client transports or the on-device host.&lt;/p&gt;

&lt;p&gt;A live portal and the screenshots used in the README are at &lt;a href="https://diiiua.github.io/ForgeMind/" rel="noopener noreferrer"&gt;diiiua.github.io/ForgeMind&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is still rough
&lt;/h2&gt;

&lt;p&gt;ForgeMind is alpha software. I would not treat it as a daily driver without watching the first tool calls yourself, especially anything that writes files, runs a shell, or drives the Accessibility service. The APK is a GitHub prerelease, not a store build.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would like to hear back
&lt;/h2&gt;

&lt;p&gt;How reliable tool calling is on the local models you actually run; whether LAN discovery finds your stack; how the Window-Tag actions behave on your OEM skin; and which parts feel unfinished for a post-alpha release.&lt;/p&gt;

&lt;p&gt;Issues and pull requests are welcome. Please do not open a PR that includes API keys, hostnames, or device identifiers.&lt;/p&gt;

&lt;h2&gt;
  
  
  Links
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;ForgeMind (Kotlin, Android, GPL-3.0, v0.2.1-alpha): &lt;a href="https://github.com/DIIIUA/ForgeMind" rel="noopener noreferrer"&gt;github.com/DIIIUA/ForgeMind&lt;/a&gt; · &lt;a href="https://github.com/DIIIUA/ForgeMind/releases/latest" rel="noopener noreferrer"&gt;Releases&lt;/a&gt; · &lt;a href="https://diiiua.github.io/ForgeMind/" rel="noopener noreferrer"&gt;Website&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I am Anton (&lt;a href="https://github.com/DIIIUA" rel="noopener noreferrer"&gt;DIIIUA&lt;/a&gt;). If you try the app, I am more interested in a failed setup or a wrong tool call than in a compliment.&lt;/p&gt;

</description>
      <category>android</category>
      <category>kotlin</category>
      <category>llm</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
