<?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: TheCodeStreet</title>
    <description>The latest articles on DEV Community by TheCodeStreet (@thecodestreet).</description>
    <link>https://dev.to/thecodestreet</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%2F1944406%2F788e81e5-324c-4a1c-a693-012f23682f15.png</url>
      <title>DEV Community: TheCodeStreet</title>
      <link>https://dev.to/thecodestreet</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/thecodestreet"/>
    <language>en</language>
    <item>
      <title>Claude Skills vs. Projects: Which One Do You Actually Need?</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Thu, 12 Nov 2026 03:30:34 +0000</pubDate>
      <link>https://dev.to/thecodestreet/claude-skills-vs-projects-which-one-do-you-actually-need-1hjb</link>
      <guid>https://dev.to/thecodestreet/claude-skills-vs-projects-which-one-do-you-actually-need-1hjb</guid>
      <description>&lt;p&gt;You've probably seen people talking about Claude Skills. And you might be thinking: wait, don't I already have Projects for this?&lt;/p&gt;

&lt;p&gt;That's a fair question. In this post I explain what Claude Skills are, how Claude decides when to use one, and exactly when you should use a Skill instead of a Project. It is the written version of my video, so you can come back to it any time.&lt;/p&gt;

&lt;h2&gt;
  
  
  🎥 Watch the Video
&lt;/h2&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/_FxutHSIbto" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;




&lt;h2&gt;
  
  
  Quick Reference
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Project&lt;/th&gt;
&lt;th&gt;Skill&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Answers&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;WHAT&lt;/strong&gt; Claude should know&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;HOW&lt;/strong&gt; Claude should do a task&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Holds&lt;/td&gt;
&lt;td&gt;Files, notes, custom instructions&lt;/td&gt;
&lt;td&gt;Step-by-step instructions, optional templates or scripts&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tied to&lt;/td&gt;
&lt;td&gt;One topic or ongoing piece of work&lt;/td&gt;
&lt;td&gt;One kind of job, across any topic&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Loaded&lt;/td&gt;
&lt;td&gt;In every chat inside the Project&lt;/td&gt;
&lt;td&gt;Only when your request matches it&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  What Are Claude Skills?
&lt;/h2&gt;

&lt;p&gt;A Skill is a set of instructions for one specific kind of job. You write it once, and Claude can use it again whenever that job comes up.&lt;/p&gt;

&lt;p&gt;Think of it like a recipe card. The card doesn't cook the meal for you. It tells the cook exactly what to do, step by step, so the meal comes out the same way every time.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Goes Inside a Skill
&lt;/h2&gt;

&lt;p&gt;A Skill lives in a folder. The main file in that folder is called &lt;code&gt;SKILL.md&lt;/code&gt;, and it has two parts:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;At the top:&lt;/strong&gt; the Skill's name and a short description of what it is for.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Below that:&lt;/strong&gt; the full instructions.
Some Skills also include extra files, like a template or a script.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Here is what a simple Skill for writing YouTube titles could look like:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;video-titles/
├── SKILL.md
├── template.md
└── check.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And the &lt;code&gt;SKILL.md&lt;/code&gt; file inside it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;---
name: video-titles
description: Write YouTube titles for my videos
---

&lt;span class="gu"&gt;## Rules
&lt;/span&gt;&lt;span class="k"&gt;- &lt;/span&gt;Keep titles under 60 characters
&lt;span class="k"&gt;- &lt;/span&gt;Use one of my 3 title styles
&lt;span class="k"&gt;- &lt;/span&gt;Run the checklist before finishing
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That small block at the top is what matters most for the next part.&lt;/p&gt;

&lt;h2&gt;
  
  
  How Claude Picks the Right Skill
&lt;/h2&gt;

&lt;p&gt;Claude doesn't read every Skill you have every time you start a chat. That would be slow, and it would fill up the conversation for no reason.&lt;/p&gt;

&lt;p&gt;Instead, it works like this:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Claude looks at the name and short description&lt;/strong&gt; of each Skill.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;When your request matches one&lt;/strong&gt;, Claude opens the full instructions and follows them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;If nothing matches&lt;/strong&gt;, Claude just carries on as normal.
&lt;strong&gt;Example:&lt;/strong&gt; you ask "Give me 5 titles for my next video". Claude scans your Skills, sees that &lt;code&gt;video-titles&lt;/code&gt; matches, and only then reads its full rules.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;This is why you can keep many Skills ready without cluttering your chats. It is also why the description matters so much. A clear description is how Claude knows when to use your Skill.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Is a Claude Project?
&lt;/h2&gt;

&lt;p&gt;A Project is a workspace for one ongoing piece of work. It holds your background: your files, your notes and your custom instructions. Every chat inside that Project can see all of it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Key Difference: WHAT vs HOW
&lt;/h2&gt;

&lt;p&gt;Here's the simple way to remember it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;A Project is about WHAT Claude should know.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A Skill is about HOW Claude should do a task.&lt;/strong&gt;
For example, a Project for your YouTube channel might hold your brand guide, your past scripts and notes about your audience. That's the background.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A Skill for writing video titles might hold your title rules, the three title styles you like and a checklist Claude runs before it finishes. That's the process.&lt;/p&gt;

&lt;h2&gt;
  
  
  3 Situations Where a Skill Is the Better Choice
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. You repeat the same process
&lt;/h3&gt;

&lt;p&gt;Maybe every time you ask for a video script, you explain the same structure again: a hook, three main points, a call to action. If you're typing that out every time, put it in a Skill.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. The same kind of job comes up in different topics
&lt;/h3&gt;

&lt;p&gt;A Project is tied to one topic. A Skill is tied to one kind of job. So if you write scripts for three different series, one script-writing Skill can help with all three.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. The instructions are long, but you only need them sometimes
&lt;/h3&gt;

&lt;p&gt;Long rules don't have to sit in every chat. A Skill loads only when your request calls for it, so your conversations stay clean.&lt;/p&gt;

&lt;h2&gt;
  
  
  When NOT to Use a Skill
&lt;/h2&gt;

&lt;p&gt;Don't turn everything into a Skill.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;One topic over many days?&lt;/strong&gt; If you need all your notes in view while the work goes on, a Project is the right tool.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Just background information?&lt;/strong&gt; Facts about your audience or details about a product belong in a Project. They aren't a step-by-step process.
A good rule: &lt;strong&gt;if it's background, use a Project. If it's a process you repeat, use a Skill.&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The Simple Test
&lt;/h2&gt;

&lt;p&gt;Ask yourself one question:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Am I explaining the same…&lt;/th&gt;
&lt;th&gt;Use&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;steps&lt;/strong&gt; over and over?&lt;/td&gt;
&lt;td&gt;✅ Skill&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;background&lt;/strong&gt; over and over?&lt;/td&gt;
&lt;td&gt;✅ Project&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  How to Find Your First Custom Skill
&lt;/h2&gt;

&lt;p&gt;The best Claude skills for you are the ones built around your own repeat work. Look back at your last few chats and find the instructions you keep typing again: a review checklist, a script structure, a formatting rule set. That is your first custom Skill.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaways
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;A Skill is a recipe card.&lt;/strong&gt; One kind of job, written once, used again and again.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude only opens a Skill when your request matches its description&lt;/strong&gt;, so you can keep many ready.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Project = WHAT, Skill = HOW.&lt;/strong&gt; Background goes in a Project, repeat processes go in a Skill.
If you like running AI tools on your own machine, I also covered &lt;a href="https://thecodestreet.dev/post/claude-code-free-vs-code-ollama-qwen/" rel="noopener noreferrer"&gt;Claude Code For Free in VS Code&lt;/a&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;What task do you repeat most with Claude? Tell me in the comments 👇 That one might be your first Skill.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://thecodestreet.dev/post/claude-skills-vs-projects/" rel="noopener noreferrer"&gt;The Code Street&lt;/a&gt;. For more AI tools and developer workflow videos, subscribe on YouTube: &lt;a href="https://www.youtube.com/@thecodestreet" rel="noopener noreferrer"&gt;@thecodestreet&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>beginners</category>
      <category>programming</category>
    </item>
    <item>
      <title>Claude Code New Update: Fable 5.2, Ultrathink &amp; 15+ Features Explained</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Wed, 11 Nov 2026 08:50:25 +0000</pubDate>
      <link>https://dev.to/thecodestreet/claude-code-new-update-fable-52-ultrathink-15-features-explained-5g59</link>
      <guid>https://dev.to/thecodestreet/claude-code-new-update-fable-52-ultrathink-15-features-explained-5g59</guid>
      <description>&lt;p&gt;Claude Code just got a new model, a way to try that model on the Pro plan this week, and a command that shows what is eating your usage limit and then fixes it for you.&lt;/p&gt;

&lt;p&gt;Those are three of the changes. This release has more than fifteen, and a few of them will change how you work every day.&lt;/p&gt;

&lt;p&gt;Here is every one of them in plain language, with an example for each.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What&lt;/th&gt;
&lt;th&gt;How to use it&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Switch to Fable 5.2&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;/model claude-fable-5-2&lt;/code&gt; (needs v2.1.321 or later)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;See and fix what is using your plan&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;/usage&lt;/code&gt;, then press &lt;code&gt;f&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Free up disk space&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;/storage&lt;/code&gt; in the desktop app&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Think harder for one turn&lt;/td&gt;
&lt;td&gt;Put &lt;code&gt;ultrathink&lt;/code&gt; in your message&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Stop auto mode blocking your own repos&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;/auto-mode-setup&lt;/code&gt; in the terminal&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;If you would rather watch than read, the full walkthrough is on YouTube:&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/nOX1v1cTq_0" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;




&lt;h2&gt;
  
  
  Fable 5.2 and the Pro Trial
&lt;/h2&gt;

&lt;p&gt;Fable 5.2 is out. It generates output up to &lt;strong&gt;30% faster&lt;/strong&gt; than Fable 5.1.&lt;/p&gt;

&lt;p&gt;The Claude Code team is honest about where it fits. Fable is a great model for big, open-ended tasks. For most everyday coding, they still recommend starting with &lt;strong&gt;Opus 5.5&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;A simple way to decide:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;"Fix this bug in my login screen" → stay on Opus 5.5.&lt;/li&gt;
&lt;li&gt;"Plan and build the whole offline sync system for my app" → this is where Fable 5.2 fits.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;To switch:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/model claude-fable-5-2
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You need Claude Code &lt;strong&gt;v2.1.321 or later&lt;/strong&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Try Fable on Pro this week
&lt;/h3&gt;

&lt;p&gt;On &lt;strong&gt;Pro&lt;/strong&gt; and &lt;strong&gt;Team Standard&lt;/strong&gt;, Fable is included for up to &lt;strong&gt;50% of your weekly limit through November 16&lt;/strong&gt;. After that, Fable uses usage credits, if you have turned them on.&lt;/p&gt;

&lt;p&gt;There may also be a Fable reset waiting for you. If you were a paid subscriber before November 9 at 10:00 AM PT, check &lt;strong&gt;Settings &amp;gt; Usage&lt;/strong&gt; on web or desktop. The deadline depends on your plan:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Plan&lt;/th&gt;
&lt;th&gt;Use the Fable reset by&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Pro, Team Standard&lt;/td&gt;
&lt;td&gt;November 16, 2026&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Max, Ultra, Team Premium&lt;/td&gt;
&lt;td&gt;March 8, 2027&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  Claude Sets Up Your Cloud Environment
&lt;/h2&gt;

&lt;p&gt;A cloud session means Claude works on your code on Anthropic's servers, not on your laptop. You can start a task from any device and it keeps going in the background.&lt;/p&gt;

&lt;p&gt;The annoying part was the setup. Dependencies, network access and API keys were all your job.&lt;/p&gt;

&lt;p&gt;On &lt;strong&gt;Pro, Max and Ultra&lt;/strong&gt;, Claude can now set up the cloud environment for you. It tells you when the session is missing a dependency, network access or an API key, and shows you the changes to approve.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Example:&lt;/strong&gt; you open your React Native project in a cloud session and ask Claude to run the tests. The tests need a package that is not installed, and they need your Firebase API key. Claude now flags both and shows you the setup changes, so the session does not just fail.&lt;/p&gt;

&lt;p&gt;Two details matter here:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;API keys go into a form&lt;/strong&gt;, so Claude does not see the key itself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The setup is saved&lt;/strong&gt;, so your next session starts with the same environment.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Free Up Disk Space with /storage
&lt;/h2&gt;

&lt;p&gt;Old sessions in the desktop app leave their worktrees on your computer. A worktree is a separate working copy of your project, and after a few weeks of daily use they add up.&lt;/p&gt;

&lt;p&gt;Type &lt;code&gt;/storage&lt;/code&gt;, or open Settings, to see how much space your sessions and worktrees take. From there you can clear the worktrees you have not used in a while. Your sessions, conversations and branches all stay. You can remove inactive tasks from the same place.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;⚠️ &lt;strong&gt;Commit first.&lt;/strong&gt; Uncommitted changes are deleted along with the worktree.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  See and Fix What Is Using Your Plan
&lt;/h2&gt;

&lt;p&gt;Running &lt;code&gt;/usage&lt;/code&gt; in the terminal already showed what was using most of your plan. It now also shows how to fix it.&lt;/p&gt;

&lt;p&gt;Press &lt;code&gt;f&lt;/code&gt;, and Claude Code sends a prompt by itself to fix whatever it can.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Example:&lt;/strong&gt; &lt;code&gt;/usage&lt;/code&gt; shows that one very long conversation is burning through most of your limit. You press &lt;code&gt;f&lt;/code&gt; and Claude takes care of it, so you do not have to work out the fix yourself.&lt;/p&gt;

&lt;h3&gt;
  
  
  Idle conversations are compacted automatically
&lt;/h3&gt;

&lt;p&gt;There is a second saving that happens on its own. If a long conversation sits idle, Claude Code now compacts it, which means it replaces the conversation with a summary, before the cache expires. That takes about one hour.&lt;/p&gt;

&lt;p&gt;Why it matters: resuming an uncached conversation means reprocessing your full context, and that costs usage. So if you go for lunch in the middle of a long session, Claude works from the summary when you come back.&lt;/p&gt;

&lt;p&gt;You will see a banner when this happens, and you can turn it off in &lt;code&gt;/config&lt;/code&gt;. It also works in VS Code and the desktop app's Code tab for conversations over 200K tokens.&lt;/p&gt;

&lt;h2&gt;
  
  
  Ultrathink Is Back
&lt;/h2&gt;

&lt;p&gt;Claude Code has effort levels, which control how hard Claude thinks before it answers. Put the word &lt;code&gt;ultrathink&lt;/code&gt; in your message and Claude uses the &lt;strong&gt;xhigh&lt;/strong&gt; effort level for that one turn only. Your session's effort setting does not change.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Example:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;ultrathink about why this migration deadlocks when two workers run it at once
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Do not put it in every message. The team says you really do not need it for most tasks. It is useful when Claude is properly stuck.&lt;/p&gt;

&lt;h2&gt;
  
  
  Teach Auto Mode About Your Setup
&lt;/h2&gt;

&lt;p&gt;Auto mode lets Claude work without asking permission at every step. By default it trusts only two things: your current working directory and that repo's own remotes.&lt;/p&gt;

&lt;p&gt;That is safe, but it means Claude gets blocked when it tries to push or send anything outside of that, even to other repos and services you own.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Example:&lt;/strong&gt; your mobile app lives in one repo and your backend API in another. Claude finishes a change and tries to push to the backend repo. It gets blocked, even though both repos are yours.&lt;/p&gt;

&lt;p&gt;The fix is one command in the terminal:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/auto-mode-setup
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It looks at your project and at the commands Claude ran in your recent sessions there. Then it drafts a list of the repos, domains and cloud storage you trust. Once you accept the list, it is saved to &lt;code&gt;~/.claude/settings.json&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  10 Smaller Updates
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Long conversations respond sooner.&lt;/strong&gt; Claude Code used to send your whole conversation with every request. For most users it now sends only the new messages.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Switching models is smoother.&lt;/strong&gt; If your conversation is too long for the new model, the model you were on summarizes it before you switch.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Queued messages survive a restart.&lt;/strong&gt; If the desktop app quits, updates or crashes, the messages you queued in a local session come back when you reopen it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cloud sessions without a repository.&lt;/strong&gt; There is a new "No repository" option on Anthropic-hosted environments, for when you do not need a repo.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cloud sessions start even if a setup script fails.&lt;/strong&gt; A failing script used to stop the session from starting.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Effort level for cloud routines.&lt;/strong&gt; The model picker now has an effort level when you create or edit a routine that runs in the cloud.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A new safety setting.&lt;/strong&gt; Turn on &lt;code&gt;permissions.blockWritesOutsideWorkingDirectories&lt;/code&gt; and the Write, Edit and NotebookEdit tools refuse paths outside your working directories in every permission mode, apart from some of Claude Code's own files. It does not limit shell commands, MCP tools or hooks.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fork subagents are on by default&lt;/strong&gt; in the desktop app and VS Code. A fork inherits your full conversation, so you do not have to explain everything again.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Background subagents use less of your limit.&lt;/strong&gt; In Anthropic's test on Pro and Max, sessions with background subagents used about 5% less.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude Code no longer installs the IDE extension for you.&lt;/strong&gt; To install or update it, use &lt;code&gt;/ide&lt;/code&gt; from an outside terminal, or your editor's extensions view.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The Claude Code team also published how they do code review. Claude reviews most of their changes before a person is asked to look, and the post explains how to set up the same thing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two Prompts from Anthropic's Engineers
&lt;/h2&gt;

&lt;p&gt;Both of these come from Anthropic's own engineering team.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;When Claude keeps stopping to ask you questions&lt;/strong&gt;, give it the go-ahead upfront:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Just build and use your best judgment. Make any call that's reversible yourself, and check in later with the calls you made.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The engineer who shared it said it unblocked a large team of agents they had running.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;When you want to test a design&lt;/strong&gt;, skip the debate about whether it is flexible enough and have Claude build on it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;In a separate worktree, do a throwaway build of [a feature we might add later] on top of this design. Don't polish it, just tell me where the design got in the way.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Example:&lt;/strong&gt; you have just designed your app's data layer. Ask for a throwaway build of offline mode on top of it. You see right away where the design gets in the way, and then you delete all of it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaways
&lt;/h2&gt;

&lt;p&gt;If you remember only three things:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Try Fable 5.2 on Pro before November 16.&lt;/strong&gt; Use it for big, open-ended tasks and keep Opus 5.5 for everyday coding.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Run &lt;code&gt;/usage&lt;/code&gt; and press &lt;code&gt;f&lt;/code&gt;&lt;/strong&gt; to cut what is eating your limit.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Run &lt;code&gt;/storage&lt;/code&gt;&lt;/strong&gt; to get your disk space back, after you commit your changes.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Everything here comes from the Claude Code team's update of November 10, 2026. The examples are my own, to make each feature easier to picture.&lt;/p&gt;




&lt;p&gt;Which update are you going to try first? Let me know in the comments.&lt;/p&gt;

&lt;p&gt;If you want to run Claude Code without a subscription, I covered that in &lt;a href="https://thecodestreet.dev/post/claude-code-free-vs-code-ollama-qwen/" rel="noopener noreferrer"&gt;Claude Code For Free in VS Code&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;I publish AI coding tutorials on &lt;a href="https://www.youtube.com/@TheCodeStreet" rel="noopener noreferrer"&gt;The Code Street&lt;/a&gt; and write them up at &lt;a href="https://thecodestreet.dev/" rel="noopener noreferrer"&gt;thecodestreet.dev&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>claude</category>
    </item>
    <item>
      <title>Liquid LFM3 Explained: Is This the Most Efficient Local AI Model?</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Thu, 05 Nov 2026 02:52:34 +0000</pubDate>
      <link>https://dev.to/thecodestreet/liquid-lfm3-explained-is-this-the-most-efficient-local-ai-model-p82</link>
      <guid>https://dev.to/thecodestreet/liquid-lfm3-explained-is-this-the-most-efficient-local-ai-model-p82</guid>
      <description>&lt;p&gt;What if a 1 GB AI model could beat a model more than five times its size?&lt;/p&gt;

&lt;p&gt;Not on everything, but on the tasks that matter most for AI agents. That's what Liquid AI is claiming with &lt;strong&gt;LFM3&lt;/strong&gt;, its new family of small, open-weight models released on &lt;strong&gt;October 27, 2026&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Here's what LFM3 is, how its architecture works, where it wins, where it doesn't, and who should actually use it.&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/S3YUCTjpX4s" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;LFM3-1.7B&lt;/strong&gt; is a 1.7B-parameter open-weight model that fits in &lt;strong&gt;~1 GB at 4-bit&lt;/strong&gt;, with a &lt;strong&gt;256K-token context window&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;It uses a &lt;strong&gt;tri-hybrid architecture&lt;/strong&gt;: 13 convolution layers, 5 sliding-window attention layers and 6 global attention layers.&lt;/li&gt;
&lt;li&gt;It &lt;strong&gt;beats Qwen 3.5 9B&lt;/strong&gt; on agent tasks like WebArena, AutomationBench and function calling.&lt;/li&gt;
&lt;li&gt;It &lt;strong&gt;loses&lt;/strong&gt; to bigger models on agentic coding (SWE-bench) and knowledge (GPQA).&lt;/li&gt;
&lt;li&gt;It's very memory-efficient: &lt;strong&gt;1.6 GB of KV cache at 128K context&lt;/strong&gt;, versus 4.3 GB for Qwen 3.5 4B.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What Is LFM3?
&lt;/h2&gt;

&lt;p&gt;LFM3 is the third generation of Liquid Foundation Models from Liquid AI. The headline model is &lt;strong&gt;LFM3-1.7B&lt;/strong&gt;, a text model built for local agents and specialised applications.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Spec&lt;/th&gt;
&lt;th&gt;LFM3-1.7B&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Parameters&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;1.7 billion&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Size at 4-bit&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;~1 GB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Context window&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;256K tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Architecture&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Tri-hybrid (convolution + attention)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Pre-training data&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;22 trillion tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Weights&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Open-weight&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;At 4-bit precision the model fits in about &lt;strong&gt;one gigabyte&lt;/strong&gt;, yet it supports a &lt;strong&gt;256,000-token context window&lt;/strong&gt;. It also outperforms Liquid's previous &lt;strong&gt;LFM2.5-2.6B&lt;/strong&gt; while using &lt;strong&gt;35% fewer parameters&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Not One Model, a Family of Six
&lt;/h2&gt;

&lt;p&gt;The first release includes six models, all built from &lt;strong&gt;LFM3-1.7B-Base&lt;/strong&gt; and all around 2B parameters:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Text:&lt;/strong&gt; for local agents and specialised apps&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Vision:&lt;/strong&gt; image understanding&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Audio:&lt;/strong&gt; speech and sound input&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Omni:&lt;/strong&gt; combines multiple input types&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Two Decision models:&lt;/strong&gt; specialised from the vision and omni models&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Liquid says every model in the family can be &lt;strong&gt;fine-tuned on a single GPU&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Is It So Fast? The Tri-Hybrid Architecture
&lt;/h2&gt;

&lt;p&gt;Pure transformers use attention in every layer, which gets expensive in memory and speed as context grows. LFM3 mixes three layer types across its 24 layers:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Layers&lt;/th&gt;
&lt;th&gt;Type&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;13&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Convolution&lt;/td&gt;
&lt;td&gt;Fast and lightweight&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;5&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Sliding-window attention&lt;/td&gt;
&lt;td&gt;Looks only at the nearest 256 tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;6&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Global attention&lt;/td&gt;
&lt;td&gt;Full attention across the whole context&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The interesting part is how Liquid found this design. They used &lt;strong&gt;hardware-in-the-loop architecture search&lt;/strong&gt;, benchmarking candidates directly on &lt;strong&gt;14 device and backend combinations&lt;/strong&gt;. These included phones (Samsung Galaxy S26 Ultra, OnePlus 15), a Raspberry Pi 5, laptops (Apple M4 Max, AMD Ryzen AI), Arm and x86 cloud CPUs, GPUs and a Qualcomm Hexagon NPU.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where It Wins: LFM3 vs Qwen 3.5 9B
&lt;/h2&gt;

&lt;p&gt;Qwen 3.5 9B is more than five times bigger. Here's where LFM3 comes out ahead:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Benchmark&lt;/th&gt;
&lt;th&gt;What it tests&lt;/th&gt;
&lt;th&gt;LFM3-1.7B&lt;/th&gt;
&lt;th&gt;Qwen 3.5 9B&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;WebArena-Verified&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Long tasks in a web browser&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;38.42&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;28.94&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;AutomationBench&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Tasks inside business apps&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;41.29&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;15.10&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;BFCL V4&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Function / tool calling&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;71.39&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;62.83&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;IFEval&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Instruction following&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;96.66&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;92.08&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;LiveCodeBench v6&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Coding problems&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;73.46&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;69.86&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;For instruction following, tool calling and agent-style tasks, LFM3 punches far above its weight. On AutomationBench it scores nearly &lt;strong&gt;3x&lt;/strong&gt; higher than Qwen 3.5 9B.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where It Loses: The Honest Comparison
&lt;/h2&gt;

&lt;p&gt;Liquid is upfront that &lt;strong&gt;larger models still win on agentic coding and knowledge&lt;/strong&gt;:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Benchmark&lt;/th&gt;
&lt;th&gt;What it tests&lt;/th&gt;
&lt;th&gt;LFM3-1.7B&lt;/th&gt;
&lt;th&gt;Qwen 3.5 9B&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;SWE-bench Verified&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Fixing real code issues&lt;/td&gt;
&lt;td&gt;43.40&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;51.20&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;GPQA Diamond&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Graduate-level science&lt;/td&gt;
&lt;td&gt;63.37&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;80.40&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;RULER 128K&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Long-context retrieval&lt;/td&gt;
&lt;td&gt;65.56&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;87.11&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A 1.7B model simply can't store as much world knowledge as a 9B one. LFM3 is a different tool for a different job, not a replacement for big models.&lt;/p&gt;

&lt;h2&gt;
  
  
  Efficiency: The Real Superpower
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;KV cache at 128K context:&lt;/strong&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;KV cache&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;LFM3-1.7B&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;1.6 GB&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Qwen 3.5 4B&lt;/td&gt;
&lt;td&gt;4.3 GB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;MiniCPM5 2B&lt;/td&gt;
&lt;td&gt;5.6 GB&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;On-device speed:&lt;/strong&gt; measured with llama.cpp at Q4_0 and a 32K context, LFM3-1.7B decodes &lt;strong&gt;1.8x to 2.8x faster&lt;/strong&gt; than the next fastest model Liquid tested, while using the least memory.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Server scale:&lt;/strong&gt; on a single NVIDIA B300 with SGLang, it can generate &lt;strong&gt;3.6 billion tokens per day&lt;/strong&gt; at 1,024 concurrent requests. Liquid also ran &lt;strong&gt;300,000+ LFM3 agents in parallel&lt;/strong&gt; to analyse the Enron email archive.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who Should Use LFM3?
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Good fit:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Older laptops or limited RAM&lt;/li&gt;
&lt;li&gt;Local AI agents that rely on tool calling and instruction following&lt;/li&gt;
&lt;li&gt;Phones and edge devices&lt;/li&gt;
&lt;li&gt;Privacy-sensitive work where data must stay on-device&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Not a good fit:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Heavy coding on large codebases&lt;/li&gt;
&lt;li&gt;Tasks that need deep factual knowledge&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For those, use a larger model like Qwen, or pair it with LFM3.&lt;/p&gt;

&lt;p&gt;AMD, Qualcomm and Ambarella all backed the launch, with plans to run LFM3 on their hardware.&lt;/p&gt;

&lt;h2&gt;
  
  
  Final Thoughts
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Verdict&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;✅ &lt;strong&gt;Tiny and fast&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;~1 GB, 256K context&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;✅ &lt;strong&gt;Great at agents and instructions&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;Beats models 5x its size on these tasks&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;⚠️ &lt;strong&gt;Not for heavy coding or deep knowledge&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;Bigger models still win there&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The direction is what excites me. Capable AI doesn't have to live in a giant data centre. It can run on your laptop.&lt;/p&gt;

&lt;p&gt;Next up, I'll install &lt;strong&gt;LFM3 locally with Ollama&lt;/strong&gt; and test it on real tasks. Follow me here or subscribe to &lt;a href="https://youtu.be/S3YUCTjpX4s" rel="noopener noreferrer"&gt;The Code Street on YouTube&lt;/a&gt; to catch it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Would you trust a 1 GB model with your daily work? Let me know in the comments.&lt;/strong&gt; 👇&lt;/p&gt;




&lt;p&gt;&lt;strong&gt;Sources&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Liquid AI: &lt;a href="https://www.liquid.ai/blog/lfm3" rel="noopener noreferrer"&gt;Introducing LFM3&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Hugging Face: &lt;a href="https://huggingface.co/LiquidAI/LFM3-1.7B" rel="noopener noreferrer"&gt;LiquidAI/LFM3-1.7B&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;All benchmark figures are as published by Liquid AI. Originally published at &lt;a href="https://thecodestreet.dev/post/liquid-lfm3-explained/" rel="noopener noreferrer"&gt;thecodestreet.dev&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>productivity</category>
      <category>programming</category>
    </item>
    <item>
      <title>Claude Code for Free in VS Code: No Login, No API Key (Ollama + Qwen 3.8)</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Mon, 02 Nov 2026 06:05:49 +0000</pubDate>
      <link>https://dev.to/thecodestreet/claude-code-for-free-in-vs-code-no-login-no-api-key-ollama-qwen-38-51f6</link>
      <guid>https://dev.to/thecodestreet/claude-code-for-free-in-vs-code-no-login-no-api-key-ollama-qwen-38-51f6</guid>
      <description>&lt;p&gt;Last time I ran Claude Code for free in the terminal. A lot of you asked the obvious follow-up: &lt;strong&gt;can I do this inside VS Code?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Yes. The official Claude Code extension can skip the Claude login entirely and talk to a model running on your own PC. No API key, no subscription, no bill. Then I gave it one prompt and it built a full Space Shooter game.&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/ffKYzsTj-3w" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;The extension normally sends your prompts to Anthropic's cloud. We point it at &lt;strong&gt;Ollama&lt;/strong&gt; instead, which runs &lt;strong&gt;Qwen 3.8&lt;/strong&gt; locally on &lt;code&gt;localhost:11434&lt;/code&gt;.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Piece&lt;/th&gt;
&lt;th&gt;Job&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;VS Code + Claude Code extension&lt;/td&gt;
&lt;td&gt;The editor and the agent you chat with&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Ollama&lt;/td&gt;
&lt;td&gt;Runs the model and serves it on port &lt;code&gt;11434&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Qwen 3.8 (27B)&lt;/td&gt;
&lt;td&gt;Actually writes the code&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;How you &lt;em&gt;use&lt;/em&gt; Claude Code doesn't change. Only where the model runs does.&lt;/p&gt;

&lt;h2&gt;
  
  
  Before you start
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;VS Code (any recent version)&lt;/li&gt;
&lt;li&gt;~32 GB of memory between RAM and VRAM&lt;/li&gt;
&lt;li&gt;~20 GB free disk (the model is ~18 GB)&lt;/li&gt;
&lt;li&gt;A GPU helps a lot (24 GB VRAM is ideal). Without one it still runs, just slower.
## 1. Install Ollama and pull the model&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Windows: grab the installer from &lt;a href="https://ollama.com" rel="noopener noreferrer"&gt;ollama.com&lt;/a&gt;. macOS / Linux:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-fsSL&lt;/span&gt; https://ollama.com/install.sh | sh
ollama &lt;span class="nt"&gt;--version&lt;/span&gt;   &lt;span class="c"&gt;# needs 0.15 or newer&lt;/span&gt;
ollama pull qwen3.8:27b
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Quick sanity check: &lt;code&gt;ollama run qwen3.8:27b&lt;/code&gt;, say hi, then &lt;code&gt;/bye&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Install the extension
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;Ctrl + Shift + X&lt;/code&gt; → search &lt;strong&gt;Claude Code&lt;/strong&gt; → make sure the publisher is &lt;strong&gt;Anthropic&lt;/strong&gt; → Install.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Set four environment variables (this skips the login)
&lt;/h2&gt;

&lt;p&gt;Open Settings (&lt;code&gt;Ctrl + ,&lt;/code&gt;), search &lt;strong&gt;Claude Code: Environment Variables&lt;/strong&gt;, click &lt;strong&gt;Edit in settings.json&lt;/strong&gt; and add:&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="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ANTHROPIC_AUTH_TOKEN"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"value"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ollama"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ANTHROPIC_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;    &lt;/span&gt;&lt;span class="nl"&gt;"value"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ollama"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ANTHROPIC_BASE_URL"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;   &lt;/span&gt;&lt;span class="nl"&gt;"value"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"http://localhost:11434"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ANTHROPIC_MODEL"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;      &lt;/span&gt;&lt;span class="nl"&gt;"value"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The two keys are placeholders. Ollama ignores them, but their presence stops the extension asking you to sign in. &lt;code&gt;ANTHROPIC_BASE_URL&lt;/code&gt; is the real trick: it sends everything to your machine instead of the cloud.&lt;/p&gt;

&lt;p&gt;Save, then &lt;code&gt;Ctrl + Shift + P&lt;/code&gt; → &lt;strong&gt;Developer: Reload Window&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Map every model to Qwen with settings.local.json
&lt;/h2&gt;

&lt;p&gt;The env vars configure the extension. This file configures &lt;strong&gt;your project&lt;/strong&gt;, so nothing ever reaches for a paid model. Create &lt;code&gt;.claude/settings.local.json&lt;/code&gt; in your project:&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="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_BASE_URL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"http://127.0.0.1:11434"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_AUTH_TOKEN"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ollama"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;""&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_DEFAULT_OPUS_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_DEFAULT_SONNET_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_DEFAULT_FABLE_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_DEFAULT_HAIKU_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"CLAUDE_CODE_SUBAGENT_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"CLAUDE_CODE_MAX_CONTEXT_TOKENS"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"65536"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The &lt;code&gt;ANTHROPIC_DEFAULT_*_MODEL&lt;/code&gt; lines cover every built-in model choice, and &lt;code&gt;CLAUDE_CODE_SUBAGENT_MODEL&lt;/code&gt; covers sub-agents.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;💡 The context here is 64k, so set Ollama's &lt;strong&gt;context length to 64k&lt;/strong&gt; too, or it will quietly forget your files mid-task. This file is machine-specific, so keep it out of Git.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  5. The test: one prompt, one game
&lt;/h2&gt;

&lt;p&gt;I opened the Claude Code panel and pasted:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Build a complete, visually stunning Space Shooter game in a single HTML file
using HTML, CSS, and JavaScript. Use Canvas for smooth gameplay, neon space
visuals, animated stars, player movement, shooting, enemy waves, particle
explosions, power-ups, score, lives, increasing difficulty, a boss fight,
and a game-over/restart screen. Make it responsive, polished, and fun to play
with keyboard controls. Keep all code in one index.html file with no external
libraries. Ensure everything works correctly.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A few minutes later: animated stars, a ship that moves and shoots, enemy waves, explosions and a boss with its own health bar. All of it generated locally, no tokens counted. (Gameplay is in the video above.)&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest part
&lt;/h2&gt;

&lt;p&gt;It's slower and less sharp than the hosted Claude models. But for side projects, learning and prototypes, a complete working game for &lt;strong&gt;$0&lt;/strong&gt; is hard to argue with.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Quick fixes if you get stuck:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Still asks you to log in:&lt;/strong&gt; re-check all four env vars, save, reload the window.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Connection refused:&lt;/strong&gt; Ollama isn't running. Open the app or run &lt;code&gt;ollama serve&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model not found:&lt;/strong&gt; the name must match &lt;code&gt;qwen3.8:27b&lt;/code&gt; exactly in both files.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  - &lt;strong&gt;Slow or forgetful:&lt;/strong&gt; set Ollama's context to 64k, close GPU-heavy apps, make sure the model fits in VRAM.
&lt;/h2&gt;

&lt;p&gt;Full guide with requirements, a variable-by-variable breakdown and an FAQ is on my blog: &lt;a href="https://thecodestreet.dev/post/claude-code-free-vs-code-ollama-qwen/" rel="noopener noreferrer"&gt;Claude Code For Free in VS Code: Full Setup Guide&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Prefer the terminal? I covered that setup here: &lt;a href="https://thecodestreet.dev/post/claude-code-for-free-ollama-qwen/" rel="noopener noreferrer"&gt;I Run Claude Code for $0 with Ollama + Qwen 3.8&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Which local model are you using for coding right now? And what should I make it build next? 👇&lt;/p&gt;

</description>
      <category>ai</category>
      <category>ollama</category>
      <category>vscode</category>
      <category>claude</category>
    </item>
    <item>
      <title>Claude Code For Free: Local Setup with Ollama + Qwen 3.8</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Thu, 29 Oct 2026 08:52:54 +0000</pubDate>
      <link>https://dev.to/thecodestreet/claude-code-for-free-local-setup-with-ollama-qwen-38-26hp</link>
      <guid>https://dev.to/thecodestreet/claude-code-for-free-local-setup-with-ollama-qwen-38-26hp</guid>
      <description>&lt;p&gt;I love Claude Code. I don't love watching the usage meter while I experiment.&lt;/p&gt;

&lt;p&gt;So I tried running it against a &lt;strong&gt;local model&lt;/strong&gt; instead. It works, it costs nothing, and the setup is shorter than this intro. Here's the TL;DR version.&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/7GaWpOj9MKs" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;h2&gt;
  
  
  The idea in one line
&lt;/h2&gt;

&lt;p&gt;Claude Code can talk to any Anthropic-compatible endpoint. Ollama (0.15+) exposes one locally, so you swap the cloud model for &lt;strong&gt;Qwen 3.8&lt;/strong&gt; running on your own GPU.&lt;/p&gt;

&lt;h2&gt;
  
  
  Before you start
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;~32 GB of memory between RAM and VRAM (the 27B model is ~18 GB)&lt;/li&gt;
&lt;li&gt;~20 GB free disk&lt;/li&gt;
&lt;li&gt;A GPU helps a lot. Without one it runs, just slowly.
## 1. Install the two tools
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Claude Code (macOS / Linux / WSL)&lt;/span&gt;
curl &lt;span class="nt"&gt;-fsSL&lt;/span&gt; https://claude.ai/install.sh | bash

&lt;span class="c"&gt;# Ollama (macOS / Linux). On Windows, grab the installer from ollama.com&lt;/span&gt;
curl &lt;span class="nt"&gt;-fsSL&lt;/span&gt; https://ollama.com/install.sh | sh

ollama &lt;span class="nt"&gt;--version&lt;/span&gt;   &lt;span class="c"&gt;# needs 0.15 or newer&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Windows users can install Claude Code from PowerShell with &lt;code&gt;irm https://claude.ai/install.ps1 | iex&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Run the one command
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;ollama launch claude &lt;span class="nt"&gt;--model&lt;/span&gt; qwen3.8:27b
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's the whole trick. Ollama starts Claude Code wired to the local model. If you'd rather download first, run &lt;code&gt;ollama pull qwen3.8:27b&lt;/code&gt; before it.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. (Optional) Make it the default for a project
&lt;/h2&gt;

&lt;p&gt;If you want plain &lt;code&gt;claude&lt;/code&gt; to always use Qwen in a repo, drop this in &lt;code&gt;.claude/settings.local.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="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_BASE_URL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"http://127.0.0.1:11434"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_AUTH_TOKEN"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ollama"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;""&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_DEFAULT_OPUS_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_DEFAULT_SONNET_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_DEFAULT_FABLE_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"ANTHROPIC_DEFAULT_HAIKU_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"CLAUDE_CODE_SUBAGENT_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"qwen3.8:27b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"CLAUDE_CODE_MAX_CONTEXT_TOKENS"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"65536"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Two things worth knowing:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The auth token is required but ignored, and the empty API key stops a real key from being picked up.&lt;/li&gt;
&lt;li&gt;Bump Ollama's context length to 64k so it matches &lt;code&gt;65536&lt;/code&gt;. Otherwise Claude Code loses track of your files mid-task.
Keep this file out of Git. It's machine-specific.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Does it actually build things?
&lt;/h2&gt;

&lt;p&gt;I gave it one prompt:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Create a simple Snake game in a single index.html file
using HTML, CSS, and vanilla JavaScript.
Use a neon cyberpunk design, arrow-key controls, score tracking,
collision detection, and a restart button.
No external libraries or files.
Save the complete, working code to index.html.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It produced a working game first try: neon styling, arrow keys, score, collisions, restart. All on my machine.&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest part
&lt;/h2&gt;

&lt;p&gt;It's slower and less sharp than the hosted Claude models. For side projects, learning, and quick throwaway builds, though, I'm not going back to paying for every experiment.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Quick fixes if you get stuck:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;command not found&lt;/code&gt;: start Ollama, then open a new terminal.&lt;/li&gt;
&lt;li&gt;Model won't load: try a smaller Qwen size.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  - Slow or forgetful: close GPU-heavy apps and raise the context length.
&lt;/h2&gt;

&lt;p&gt;Full step-by-step guide (with troubleshooting and FAQ) is on my blog: &lt;a href="https://thecodestreet.dev/post/claude-code-for-free-ollama-qwen/" rel="noopener noreferrer"&gt;Claude Code For Free: Local Setup with Ollama + Qwen 3.8&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Next up: the same free setup inside VS Code. Have you tried running coding agents on local models? What's your go-to model? 👇&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>ollama</category>
      <category>llm</category>
    </item>
    <item>
      <title>How to Use Claude Code with Opus 5.5 &amp; Fable 5.1 for Free ( No Card Required)</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Mon, 26 Oct 2026 10:26:54 +0000</pubDate>
      <link>https://dev.to/thecodestreet/how-to-use-claude-code-with-opus-55-fable-51-for-free-no-card-required-4ngi</link>
      <guid>https://dev.to/thecodestreet/how-to-use-claude-code-with-opus-55-fable-51-for-free-no-card-required-4ngi</guid>
      <description>&lt;p&gt;Want to use &lt;strong&gt;Claude Code&lt;/strong&gt; with top models like &lt;strong&gt;Claude Opus 5.5&lt;/strong&gt; and &lt;strong&gt;Fable 5.1&lt;/strong&gt;, but don't want to pay for a subscription or fund an API account yet?&lt;/p&gt;

&lt;p&gt;In this guide, I'll show you how to connect Claude Code to a third-party AI gateway that gives you &lt;strong&gt;500,000 free tokens every month&lt;/strong&gt;, with &lt;strong&gt;no credit card required&lt;/strong&gt;. The setup takes about five minutes.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;⚠️ &lt;strong&gt;Before you start:&lt;/strong&gt; This uses a third-party service that is &lt;strong&gt;not affiliated with or endorsed by Anthropic&lt;/strong&gt;. Free plan limits, available models, and pricing can change at any time. Read the disclaimer at the end before you use it for anything serious.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  🎥 Watch the Video Tutorial
&lt;/h2&gt;

&lt;p&gt;Prefer watching? Here's the full step-by-step walkthrough:&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/W5clGOuOySI" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;h2&gt;
  
  
  What You Get
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Feature&lt;/th&gt;
&lt;th&gt;Details&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Free tokens&lt;/td&gt;
&lt;td&gt;500,000 per month&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Credit card&lt;/td&gt;
&lt;td&gt;Not required&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Models&lt;/td&gt;
&lt;td&gt;Claude Opus 5.5, Fable 5.1, and more&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Works with&lt;/td&gt;
&lt;td&gt;Claude Code (via third-party inference)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Best for&lt;/td&gt;
&lt;td&gt;Learning, side projects, prototyping&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  How much is 500K tokens?
&lt;/h3&gt;

&lt;p&gt;A &lt;strong&gt;token&lt;/strong&gt; is a small chunk of text, often part of a word. A common rule of thumb for English:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;1 token ≈ 4 characters ≈ ¾ of a word&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;1,000 tokens ≈ 750 words&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;500,000 tokens ≈ 375,000 words&lt;/strong&gt;
Input &lt;em&gt;and&lt;/em&gt; output both count. Coding sessions also send file contents and conversation history as context, so a long agentic session uses more than you might expect. Still, 500K a month is plenty for learning and for several small projects.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why Use Claude Code with Free Tokens?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Students and beginners&lt;/strong&gt; can learn AI-assisted coding at zero cost.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Indie developers&lt;/strong&gt; can prototype ideas without upfront spend.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Curious developers&lt;/strong&gt; can try premium models like Opus 5.5 before committing to a paid plan.&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;Content creators and educators&lt;/strong&gt; can demo Claude Code without burning a personal budget.&lt;/p&gt;
&lt;h2&gt;
  
  
  Prerequisites
&lt;/h2&gt;
&lt;/li&gt;
&lt;li&gt;&lt;p&gt;A computer running Windows, macOS, or Linux&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;An internet connection&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;A free account on the AI gateway: &lt;a href="https://radoxai.com" rel="noopener noreferrer"&gt;Radox AI&lt;/a&gt;&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;About 5 minutes&lt;/p&gt;
&lt;h2&gt;
  
  
  Step-by-Step Setup
&lt;/h2&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Step 1: Install Claude Code
&lt;/h3&gt;

&lt;p&gt;If you don't already have Claude Code installed, follow the official installation guide:&lt;/p&gt;

&lt;p&gt;👉 &lt;a href="https://docs.claude.com/en/docs/claude-code/overview" rel="noopener noreferrer"&gt;Claude Code documentation&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Once it's installed, open it once to make sure it launches correctly.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 2: Enable Developer Mode
&lt;/h3&gt;

&lt;p&gt;Developer Mode unlocks the advanced configuration options, including &lt;strong&gt;third-party inference&lt;/strong&gt;. That's what lets Claude Code send requests to a provider other than the default.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Open the settings.&lt;/li&gt;
&lt;li&gt;Find and turn on &lt;strong&gt;Developer Mode&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Open the &lt;strong&gt;third-party inference&lt;/strong&gt; configuration.
&amp;gt; 💡 Menu names can change between versions. If you can't find an option, check the video above or the latest official docs.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  Step 3: Add Your API Key and Endpoint
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;Sign up on &lt;a href="https://radoxai.com" rel="noopener noreferrer"&gt;Radox AI&lt;/a&gt;. No card is needed.&lt;/li&gt;
&lt;li&gt;Open your dashboard and &lt;strong&gt;create an API key&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Copy the &lt;strong&gt;Gateway Base URL (endpoint)&lt;/strong&gt; from the dashboard.&lt;/li&gt;
&lt;li&gt;Paste both into the third-party inference settings in Claude Code.
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;API Key:   &amp;lt;your-api-key&amp;gt;
Endpoint:  &amp;lt;gateway-base-url-from-your-dashboard&amp;gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;blockquote&gt;
&lt;p&gt;🔐 &lt;strong&gt;Treat your API key like a password.&lt;/strong&gt; Never commit it to Git, paste it in public chats, or show it on screen in recordings.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h3&gt;
  
  
  Step 4: Test, Reload, and Launch
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;Click &lt;strong&gt;Test Connection&lt;/strong&gt; and wait for a success message.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reload&lt;/strong&gt; and &lt;strong&gt;launch&lt;/strong&gt; Claude Code.&lt;/li&gt;
&lt;li&gt;Select &lt;strong&gt;Claude Opus 5.5&lt;/strong&gt; or &lt;strong&gt;Fable 5.1&lt;/strong&gt; as your model.
You're ready to code with free tokens. 🎉&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Try It: Your First Prompt
&lt;/h2&gt;

&lt;p&gt;Here's a fun prompt to test the setup. It builds a complete, polished browser game in one shot:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Build a single-file HTML/CSS/JS game called "Orbit Rush". A glowing orb
circles a central star. Tap or press Space to switch between an inner and
outer orbit (with a smooth transition). Meteors fly in from the edges, so
dodge them and collect energy shards for points. Speed ramps up over time.

Visuals matter most: dark space background with parallax stars, neon glow,
a fading comet trail behind the orb, particle bursts on collecting and
crashing, a subtle screen shake on game over, and a polished title screen,
score HUD, and restart button. Add small WebAudio sound effects (no external
files) and save the high score in localStorage. Use canvas, 60fps, responsive
to any screen size, no libraries. Make it feel like a finished game.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open the generated HTML file in your browser and play. A prompt like this is a good way to see what Opus 5.5 can do with a single request.&lt;/p&gt;

&lt;h2&gt;
  
  
  Tips to Make Your Free Tokens Last Longer
&lt;/h2&gt;

&lt;p&gt;Every message sends context to the model, so a little care stretches your 500K much further:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Start fresh for new tasks.&lt;/strong&gt; Use &lt;code&gt;/clear&lt;/code&gt; when you switch to something unrelated, so old conversation history isn't re-sent.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Compact long sessions.&lt;/strong&gt; Use &lt;code&gt;/compact&lt;/code&gt; to summarize a long conversation and cut the context size.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Be specific.&lt;/strong&gt; A clear, detailed prompt avoids back-and-forth that wastes tokens.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Point to the exact files.&lt;/strong&gt; Name the files that matter instead of letting the agent explore your whole repo.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Match the model to the task.&lt;/strong&gt; Save the most powerful models for hard problems, and use lighter ones for simple edits if your gateway offers them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Check your usage.&lt;/strong&gt; Keep an eye on the token counter in your gateway dashboard so you don't run out mid-project.
## Troubleshooting&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;Test Connection fails&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Check that the API key was copied fully, with no extra spaces.&lt;/li&gt;
&lt;li&gt;Check that the endpoint URL matches the dashboard exactly, including &lt;code&gt;https://&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Make sure your free quota hasn't run out for the month.
&lt;strong&gt;The model isn't listed&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;Reload Claude Code after saving the settings.&lt;/li&gt;
&lt;li&gt;Check your gateway dashboard to see which models your plan currently includes.
&lt;strong&gt;Responses suddenly stop&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;You may have hit your monthly limit. Check usage in the dashboard.
## FAQ&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Is this really free?&lt;/strong&gt;&lt;br&gt;
At the time of writing, the gateway offers 500,000 tokens per month with no credit card. Free plans can change at any time, so check the current terms on the site.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is this official from Anthropic?&lt;/strong&gt;&lt;br&gt;
No. It's a third-party gateway that Claude Code connects to through its third-party inference support. It isn't affiliated with or endorsed by Anthropic.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can I use this for work projects?&lt;/strong&gt;&lt;br&gt;
I wouldn't recommend it for sensitive or proprietary code. Your prompts and code pass through a third-party service. For professional work, use an official plan or the official API.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What happens when I run out of tokens?&lt;/strong&gt;&lt;br&gt;
Requests will fail until your quota resets next month, or until you switch to another provider or plan.&lt;/p&gt;

&lt;h2&gt;
  
  
  Disclaimer
&lt;/h2&gt;

&lt;p&gt;This guide is for &lt;strong&gt;educational purposes&lt;/strong&gt;. The gateway used here is a third-party service, &lt;strong&gt;not affiliated with Anthropic&lt;/strong&gt;. Free plan limits, model availability, and terms &lt;strong&gt;can change at any time without notice&lt;/strong&gt;. Always:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Read the provider's terms and privacy policy&lt;/li&gt;
&lt;li&gt;Keep your API key private&lt;/li&gt;
&lt;li&gt;Avoid sending sensitive, personal, or proprietary data through third-party services
## Wrapping Up&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;With a few minutes of setup, you can use Claude Code with Opus 5.5 and Fable 5.1 on &lt;strong&gt;500K free tokens a month&lt;/strong&gt;. That's a great way to learn AI-assisted development, prototype ideas, and try premium models before paying for anything.&lt;/p&gt;

&lt;p&gt;If this helped, &lt;strong&gt;watch the full video tutorial&lt;/strong&gt; above and &lt;strong&gt;subscribe to The Code Street&lt;/strong&gt; on YouTube for more indie dev and AI coding content.&lt;/p&gt;

&lt;p&gt;💬 &lt;strong&gt;Are you using Claude Code yet? Paid plan, API, or free tokens? Tell me in the comments!&lt;/strong&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>free</category>
      <category>beginners</category>
    </item>
    <item>
      <title>REA GitHub Tutorial: Reverse Engineer Anything with Claude Code (Full Setup + Fixes)</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Tue, 20 Oct 2026 19:28:48 +0000</pubDate>
      <link>https://dev.to/thecodestreet/rea-github-tutorial-reverse-engineer-anything-with-claude-code-full-setup-fixes-1g89</link>
      <guid>https://dev.to/thecodestreet/rea-github-tutorial-reverse-engineer-anything-with-claude-code-full-setup-fixes-1g89</guid>
      <description>&lt;p&gt;&lt;strong&gt;REA (Reverse Engineer Anything)&lt;/strong&gt; is the viral GitHub repo that gives AI coding agents like &lt;strong&gt;Claude Code&lt;/strong&gt; the tools to look inside an installed app — without the source code. I installed it on Windows, pointed it at &lt;strong&gt;VS Code&lt;/strong&gt;, and wrote down every error I hit along the way.&lt;/p&gt;

&lt;p&gt;🎥 &lt;strong&gt;Prefer video?&lt;/strong&gt; Watch the full walkthrough here: [&lt;a href="https://youtu.be/H_KYsLCeWUE" rel="noopener noreferrer"&gt;Walkthrough&lt;/a&gt;]&lt;/p&gt;




&lt;h2&gt;
  
  
  What is REA?
&lt;/h2&gt;

&lt;p&gt;REA is an &lt;strong&gt;MCP server&lt;/strong&gt; plus a set of workflow instructions (a "skill") for your AI agent. Instead of guessing how an app works from screenshots, the agent can call REA's analysis tools — 207 of them in version 6.7.0 — to inspect the real files.&lt;/p&gt;

&lt;p&gt;Think of it like a restaurant:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;You&lt;/strong&gt; are the customer — you ask a question in plain English.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude Code&lt;/strong&gt; is the chef — it decides what to do.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;REA&lt;/strong&gt; is the chef's toolbox.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The app&lt;/strong&gt; is the ingredient being inspected.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It supports JavaScript and Electron apps out of the box. Native &lt;code&gt;.exe&lt;/code&gt;/&lt;code&gt;.dll&lt;/code&gt; analysis needs Ghidra, IDA or Hopper installed separately, and full Android APK analysis needs a Java JDK.&lt;/p&gt;

&lt;h3&gt;
  
  
  Hype vs reality
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;✅ It finds real evidence inside an app's files&lt;/li&gt;
&lt;li&gt;✅ It explains that evidence in plain language&lt;/li&gt;
&lt;li&gt;✅ It tells you what it &lt;strong&gt;couldn't&lt;/strong&gt; figure out&lt;/li&gt;
&lt;li&gt;❌ It is &lt;strong&gt;not&lt;/strong&gt; a one-click "clone this app" button&lt;/li&gt;
&lt;li&gt;❌ Heavily protected apps need extra tools and may still leave open questions&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;My favourite part: findings come with a confidence level. In my test, the Electron fuse check came back as &lt;code&gt;"confidence": "observed"&lt;/code&gt; from the shipped artifact, with the file's SHA-256 attached as proof.&lt;/p&gt;




&lt;h2&gt;
  
  
  How to install REA
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Prerequisite:&lt;/strong&gt; Node.js 22.19+, 24.11+ or 26+.&lt;/p&gt;

&lt;h3&gt;
  
  
  Mac &amp;amp; Linux (guided)
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx rea-agents setup
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The wizard asks which agent you use and registers REA for you.&lt;/p&gt;

&lt;h3&gt;
  
  
  Windows (manual — this is what worked for me)
&lt;/h3&gt;

&lt;p&gt;On Windows, setup deliberately doesn't edit your agent's config, so register it yourself:&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;node&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;-v&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;npm&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;install&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;-g&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;rea-agents&lt;/span&gt;&lt;span class="err"&gt;@&lt;/span&gt;&lt;span class="nx"&gt;6.7.0&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;rea&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;doctor&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="n"&gt;claude&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;mcp&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;add&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nt"&gt;--scope&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;user&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;rea&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="o"&gt;--&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;cmd&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;/c&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;rea&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nx"&gt;mcp&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then &lt;strong&gt;fully restart Claude Code&lt;/strong&gt; and run &lt;code&gt;/mcp&lt;/code&gt; — you should see &lt;code&gt;rea · connected&lt;/code&gt;.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;code&gt;cmd /c&lt;/code&gt; matters on Windows: &lt;code&gt;rea&lt;/code&gt; is a &lt;code&gt;.cmd&lt;/code&gt; launcher that Claude Code can't start directly.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;⚠️ Install only from the official repo — &lt;a href="https://github.com/morluto/rea" rel="noopener noreferrer"&gt;github.com/morluto/rea&lt;/a&gt; — or the npm package &lt;a href="https://www.npmjs.com/package/rea-agents" rel="noopener noreferrer"&gt;&lt;code&gt;rea-agents&lt;/code&gt;&lt;/a&gt;. Copycat repos are a real risk with tools like this.&lt;/p&gt;




&lt;h2&gt;
  
  
  Demo: reverse engineering VS Code
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu819g07gaw7g51zpzibu.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu819g07gaw7g51zpzibu.png" alt="REA GitHub Walkthrough" width="800" height="533"&gt;&lt;/a&gt;&lt;br&gt;
I gave Claude Code this prompt (static analysis only — nothing launched or modified):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Use the rea MCP server to statically analyse VS Code at
%LOCALAPPDATA%\Programs\Microsoft VS Code. Do not launch or modify it.
1. inspect_pe_resources and inspect_electron_fuses on Code.exe
2. Read resources\app\product.json and list every server it talks to
3. analyze_javascript_application on resources\app and map its processes
Explain each finding in plain English.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  1. What Code.exe really is
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Component&lt;/th&gt;
&lt;th&gt;Version&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;VS Code&lt;/td&gt;
&lt;td&gt;1.142.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Electron&lt;/td&gt;
&lt;td&gt;43.7.7&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Chromium&lt;/td&gt;
&lt;td&gt;150.0.7871.250&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Node.js&lt;/td&gt;
&lt;td&gt;24.21.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Size&lt;/td&gt;
&lt;td&gt;238 MB, 64-bit, signed by Microsoft&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The internal name? &lt;code&gt;electron.exe&lt;/code&gt;. VS Code is a web app in a desktop wrapper.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Nine Electron security fuses
&lt;/h3&gt;

&lt;p&gt;Fuses are on/off switches baked into the executable at build time. REA read all nine. Highlights:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;On:&lt;/strong&gt; &lt;code&gt;RunAsNode&lt;/code&gt;, &lt;code&gt;EnableNodeOptionsEnvironmentVariable&lt;/code&gt;, &lt;code&gt;EnableNodeCliInspectArguments&lt;/code&gt; — VS Code needs these for its CLI and helper processes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Off:&lt;/strong&gt; &lt;code&gt;EnableEmbeddedAsarIntegrityValidation&lt;/code&gt;, &lt;code&gt;OnlyLoadAppFromAsar&lt;/code&gt; — VS Code checks its own files with checksums in &lt;code&gt;product.json&lt;/code&gt; instead.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These are design trade-offs, not bugs — but they're exactly why Electron apps are interesting to study.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Who VS Code talks to
&lt;/h3&gt;

&lt;p&gt;From &lt;code&gt;product.json&lt;/code&gt;: the update server, the Extension Marketplace, a GitHub-hosted MCP server gallery, Settings Sync, an A/B experimentation service, Bing-powered settings search, a voice service, and a CDN hosting AI agent SDKs. Telemetry is on by default, with an opt-out.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Seven processes
&lt;/h3&gt;

&lt;p&gt;Main process, workbench window, shared process, extension host, pty host (terminal), file watcher, and a newer agent host. That's why one frozen extension doesn't freeze your editor — extensions run in their own process.&lt;/p&gt;




&lt;h2&gt;
  
  
  Troubleshooting: every error I hit
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Error&lt;/th&gt;
&lt;th&gt;Cause&lt;/th&gt;
&lt;th&gt;Fix&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;'claude' is not recognized&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;claude.exe&lt;/code&gt; is in &lt;code&gt;%USERPROFILE%\.local\bin&lt;/code&gt;, which wasn't on PATH&lt;/td&gt;
&lt;td&gt;Add that folder to your user PATH, open a new terminal&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;MCP server rea already exists in user config&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Already registered&lt;/td&gt;
&lt;td&gt;Nothing to fix — restart Claude Code&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;rea&lt;/code&gt; missing from &lt;code&gt;/mcp&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;Not restarted, or a different session/PC&lt;/td&gt;
&lt;td&gt;Fully close and reopen Claude Code on the same PC&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;provider_unavailable&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;No Ghidra/IDA/Hopper found&lt;/td&gt;
&lt;td&gt;Install Ghidra + JDK 21, set &lt;code&gt;GHIDRA_INSTALL_DIR&lt;/code&gt;, run &lt;code&gt;rea doctor&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Server won't start on Windows&lt;/td&gt;
&lt;td&gt;Missing &lt;code&gt;cmd /c&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;&lt;code&gt;claude mcp add --scope user rea -- cmd /c rea mcp&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;APK analysis fails&lt;/td&gt;
&lt;td&gt;JRE instead of JDK&lt;/td&gt;
&lt;td&gt;Install a full Java JDK&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;Pro tip:&lt;/strong&gt; when anything breaks, run &lt;code&gt;rea doctor&lt;/code&gt; first. It's read-only and tells you exactly what's missing.&lt;/p&gt;




&lt;h2&gt;
  
  
  My verdict
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;For understanding apps — yes.&lt;/strong&gt; What used to take hours in a disassembler took minutes, in plain English.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not a magic clone button.&lt;/strong&gt; It gives your agent evidence; building is still your job.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Always verify&lt;/strong&gt; anything marked "inferred" or "unknown".&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;As an app developer, I'll use it to check my own release builds (what really ships in my APK), learn how good open-source apps solve a feature, and audit what the tools I install connect to.&lt;/p&gt;

&lt;h2&gt;
  
  
  Use it responsibly
&lt;/h2&gt;

&lt;p&gt;From REA's README: it's for &lt;em&gt;lawful research, education, interoperability, software preservation&lt;/em&gt; — and software you own or are authorized to examine. Learn from apps; don't copy paid ones.&lt;/p&gt;




&lt;p&gt;&lt;strong&gt;Watch the full video:&lt;/strong&gt; &lt;a href="https://youtu.be/H_KYsLCeWUE" rel="noopener noreferrer"&gt;Full Detailed Video&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;What should I reverse engineer next? Drop it in the comments 👇&lt;/p&gt;

&lt;p&gt;&lt;em&gt;I'm Rajiv — I make videos about AI tools that actually work on &lt;a href="https://www.youtube.com/@thecodestreet" rel="noopener noreferrer"&gt;The Code Street&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>claude</category>
      <category>beginners</category>
    </item>
    <item>
      <title>How to Use Qwen 3 with VS Code (FREE) | Complete Setup with Ollama</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Tue, 04 Aug 2026 17:31:59 +0000</pubDate>
      <link>https://dev.to/thecodestreet/how-to-use-qwen-3-with-vs-code-free-complete-setup-with-ollama-29f8</link>
      <guid>https://dev.to/thecodestreet/how-to-use-qwen-3-with-vs-code-free-complete-setup-with-ollama-29f8</guid>
      <description>&lt;p&gt;Have you ever wanted an AI coding assistant inside &lt;strong&gt;VS Code&lt;/strong&gt; without paying for GitHub Copilot or other monthly subscriptions?&lt;/p&gt;

&lt;p&gt;The good news is—you can.&lt;/p&gt;

&lt;p&gt;Using &lt;strong&gt;Qwen 3&lt;/strong&gt; and &lt;strong&gt;Ollama&lt;/strong&gt;, you can run a powerful coding model directly on your computer and use it inside VS Code.&lt;/p&gt;

&lt;p&gt;No API costs.&lt;/p&gt;

&lt;p&gt;No cloud dependency.&lt;/p&gt;

&lt;p&gt;Your code stays on your machine.&lt;/p&gt;




&lt;h1&gt;
  
  
  📺 Watch the Complete Video Tutorial
&lt;/h1&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;🎥 YouTube Video:&lt;/strong&gt;&lt;br&gt;&lt;br&gt;
&lt;strong&gt;How to Use Qwen 3 with VS Code (FREE) | Complete Setup with Ollama&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/1TUeKUC2ffk"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;




&lt;h1&gt;
  
  
  What You'll Learn
&lt;/h1&gt;

&lt;p&gt;By the end of this tutorial you'll know how to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Install Ollama&lt;/li&gt;
&lt;li&gt;Download Qwen 3&lt;/li&gt;
&lt;li&gt;Connect Qwen 3 with VS Code&lt;/li&gt;
&lt;li&gt;Configure OpenCode&lt;/li&gt;
&lt;li&gt;Generate your first AI code&lt;/li&gt;
&lt;li&gt;Fix common setup issues&lt;/li&gt;
&lt;li&gt;Use local AI effectively&lt;/li&gt;
&lt;/ul&gt;




&lt;h1&gt;
  
  
  What is Qwen 3?
&lt;/h1&gt;

&lt;p&gt;Qwen 3 is Alibaba's latest family of open-source large language models.&lt;/p&gt;

&lt;p&gt;It performs well for:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Code Generation&lt;/li&gt;
&lt;li&gt;Debugging&lt;/li&gt;
&lt;li&gt;Code Explanation&lt;/li&gt;
&lt;li&gt;Refactoring&lt;/li&gt;
&lt;li&gt;General Programming Tasks&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Since it can run locally using Ollama, you don't need to rely on cloud APIs for everyday coding assistance.&lt;/p&gt;




&lt;h1&gt;
  
  
  What is Ollama?
&lt;/h1&gt;

&lt;p&gt;Ollama lets you run Large Language Models directly on your computer.&lt;/p&gt;

&lt;p&gt;Benefits include:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Privacy&lt;/li&gt;
&lt;li&gt;No API Charges&lt;/li&gt;
&lt;li&gt;Offline Support&lt;/li&gt;
&lt;li&gt;Fast Local Responses&lt;/li&gt;
&lt;li&gt;Easy Model Management&lt;/li&gt;
&lt;/ul&gt;




&lt;h1&gt;
  
  
  Step 1 — Install Ollama
&lt;/h1&gt;

&lt;p&gt;Download Ollama from:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://ollama.com" rel="noopener noreferrer"&gt;https://ollama.com&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Verify installation:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;ollama &lt;span class="nt"&gt;--version&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h1&gt;
  
  
  Step 2 — Download Qwen 3
&lt;/h1&gt;

&lt;p&gt;Run:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;ollama pull qwen3.6:latest
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Verify:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;ollama list
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You should now see the Qwen 3 model installed locally.&lt;/p&gt;




&lt;h1&gt;
  
  
  Step 4 — Connect Qwen 3 with VS Code
&lt;/h1&gt;

&lt;p&gt;Configure OpenCode to use your Ollama server.&lt;/p&gt;

&lt;p&gt;Typical endpoint:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;http://localhost:11434
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;or&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;http://localhost:11434/v1
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;depending on your configuration.&lt;/p&gt;




&lt;h1&gt;
  
  
  Step 5 — Generate Your First Code
&lt;/h1&gt;

&lt;p&gt;Instead of asking:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Build an entire React application&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Start with something simple.&lt;/p&gt;

&lt;p&gt;Example:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Create a reusable React Button component using functional components and CSS Modules.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Small prompts generally produce better results with local models.&lt;/p&gt;




&lt;h1&gt;
  
  
  Common Mistakes
&lt;/h1&gt;

&lt;h3&gt;
  
  
  ❌ Asking for an entire application
&lt;/h3&gt;

&lt;p&gt;Instead:&lt;/p&gt;

&lt;p&gt;Break the project into smaller components.&lt;/p&gt;




&lt;h3&gt;
  
  
  ❌ Choosing the wrong model
&lt;/h3&gt;

&lt;p&gt;Use a coding-capable Qwen model whenever possible.&lt;/p&gt;




&lt;h3&gt;
  
  
  ❌ Expecting cloud-model behavior
&lt;/h3&gt;

&lt;p&gt;Local models work best when you:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Keep prompts focused&lt;/li&gt;
&lt;li&gt;Iterate gradually&lt;/li&gt;
&lt;li&gt;Review generated code&lt;/li&gt;
&lt;/ul&gt;




&lt;h1&gt;
  
  
  My Experience
&lt;/h1&gt;

&lt;p&gt;After experimenting with Qwen 3, I realized that the biggest limitation wasn't the model itself—it was my prompting strategy.&lt;/p&gt;

&lt;p&gt;Once I started building projects step by step instead of expecting an entire application in one response, the quality improved significantly.&lt;/p&gt;




&lt;h1&gt;
  
  
  What's Next?
&lt;/h1&gt;

&lt;p&gt;This article is Part 1 of the series.&lt;/p&gt;

&lt;p&gt;In Part 2 we'll build a complete &lt;strong&gt;GitHub Profile Finder&lt;/strong&gt; using:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;React&lt;/li&gt;
&lt;li&gt;GitHub API&lt;/li&gt;
&lt;li&gt;Qwen 3&lt;/li&gt;
&lt;li&gt;VS Code&lt;/li&gt;
&lt;li&gt;Ollama&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;…all with step-by-step prompts.&lt;/p&gt;




&lt;h1&gt;
  
  
  Resources
&lt;/h1&gt;

&lt;h2&gt;
  
  
  Ollama
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://ollama.com" rel="noopener noreferrer"&gt;https://ollama.com&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Qwen Models
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://ollama.com/library/qwen3" rel="noopener noreferrer"&gt;https://ollama.com/library/qwen3&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  VS Code
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://code.visualstudio.com/" rel="noopener noreferrer"&gt;https://code.visualstudio.com/&lt;/a&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  Conclusion
&lt;/h1&gt;

&lt;p&gt;Running Qwen 3 locally inside VS Code is a great option if you want a private, free AI coding assistant.&lt;/p&gt;

&lt;p&gt;Whether you're learning React, JavaScript, or working on professional software projects, local AI models have become capable enough to be part of your daily development workflow.&lt;/p&gt;




&lt;h2&gt;
  
  
  🎥 Prefer Watching?
&lt;/h2&gt;

&lt;p&gt;If you found this article helpful, consider following &lt;strong&gt;TheCodeStreet&lt;/strong&gt; for more tutorials on:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;AI&lt;/li&gt;
&lt;li&gt;.NET&lt;/li&gt;
&lt;li&gt;C#&lt;/li&gt;
&lt;li&gt;Ollama&lt;/li&gt;
&lt;li&gt;Qwen&lt;/li&gt;
&lt;li&gt;AI Agents&lt;/li&gt;
&lt;li&gt;Semantic Kernel&lt;/li&gt;
&lt;li&gt;VS Code&lt;/li&gt;
&lt;li&gt;Software Development&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Happy Coding! 🚀&lt;/p&gt;

</description>
      <category>ai</category>
      <category>qwen</category>
      <category>programming</category>
      <category>beginners</category>
    </item>
    <item>
      <title>🚨 BREAKING: Microsoft Pauses GitHub Copilot Subscriptions — What Developers Need to Know</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Fri, 24 Apr 2026 04:43:31 +0000</pubDate>
      <link>https://dev.to/thecodestreet/breaking-microsoft-pauses-github-copilot-subscriptions-what-developers-need-to-know-5oo</link>
      <guid>https://dev.to/thecodestreet/breaking-microsoft-pauses-github-copilot-subscriptions-what-developers-need-to-know-5oo</guid>
      <description>&lt;p&gt;GitHub Copilot has become a daily coding companion for many of us…&lt;br&gt;
And suddenly, new subscriptions are paused.&lt;/p&gt;

&lt;p&gt;If you’re using AI tools for coding, this update might have surprised you.&lt;/p&gt;

&lt;p&gt;👉 &lt;strong&gt;Microsoft has paused new GitHub Copilot subscriptions.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Yes — the tool many developers rely on daily just hit an unexpected pause.&lt;/p&gt;

&lt;p&gt;But the real question is: &lt;strong&gt;Why? And what does this mean for us?&lt;/strong&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  🎥 Watch Full Breakdown (Quick Video)
&lt;/h2&gt;

&lt;p&gt;Before we dive deep, I’ve explained everything in this video:&lt;/p&gt;

&lt;p&gt;👉 &lt;a href="https://youtu.be/YvtdyQfV8p4" rel="noopener noreferrer"&gt;https://youtu.be/YvtdyQfV8p4&lt;/a&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  🤯 What Happened?
&lt;/h2&gt;

&lt;p&gt;Microsoft has &lt;strong&gt;temporarily paused new GitHub Copilot subscriptions&lt;/strong&gt;, creating confusion across the developer community.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Existing users? ✅ Still active
&lt;/li&gt;
&lt;li&gt;New subscriptions? ❌ Currently paused
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This is not a shutdown — but definitely something worth paying attention to.&lt;/p&gt;




&lt;h2&gt;
  
  
  🧠 Why Did Microsoft Do This? (Possible Reasons)
&lt;/h2&gt;

&lt;p&gt;While there’s no single official reason publicly confirmed in detail, here are some strong possibilities:&lt;/p&gt;

&lt;h3&gt;
  
  
  1️⃣ Infrastructure &amp;amp; Scaling Challenges
&lt;/h3&gt;

&lt;p&gt;AI coding tools like Copilot require massive compute power.&lt;br&gt;&lt;br&gt;
👉 Sudden demand spikes can force companies to &lt;strong&gt;pause onboarding&lt;/strong&gt;.&lt;/p&gt;




&lt;h3&gt;
  
  
  2️⃣ Strategic Shift in AI Offerings
&lt;/h3&gt;

&lt;p&gt;Microsoft is heavily investing in AI across products like:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Azure AI
&lt;/li&gt;
&lt;li&gt;Developer tools
&lt;/li&gt;
&lt;li&gt;Enterprise AI solutions
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;👉 This pause could be part of a &lt;strong&gt;bigger product restructuring&lt;/strong&gt;.&lt;/p&gt;




&lt;h3&gt;
  
  
  3️⃣ Pricing / Model Changes Incoming
&lt;/h3&gt;

&lt;p&gt;We’ve seen this pattern before in AI tools.&lt;/p&gt;

&lt;p&gt;👉 Pause → Update → Relaunch with new pricing/features&lt;/p&gt;




&lt;h3&gt;
  
  
  4️⃣ Quality &amp;amp; Performance Optimization
&lt;/h3&gt;

&lt;p&gt;Microsoft might be improving:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Response quality
&lt;/li&gt;
&lt;li&gt;Latency
&lt;/li&gt;
&lt;li&gt;Model upgrades
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;👉 Better experience before scaling further.&lt;/p&gt;




&lt;h2&gt;
  
  
  👨‍💻 What This Means for Developers
&lt;/h2&gt;

&lt;p&gt;This impacts different groups differently:&lt;/p&gt;

&lt;h3&gt;
  
  
  🔹 Existing Users
&lt;/h3&gt;

&lt;p&gt;No major impact (for now)&lt;br&gt;&lt;br&gt;
👉 You can continue using Copilot normally&lt;/p&gt;




&lt;h3&gt;
  
  
  🔹 New Users
&lt;/h3&gt;

&lt;p&gt;You might not be able to subscribe immediately&lt;br&gt;&lt;br&gt;
👉 Need to wait for updates&lt;/p&gt;




&lt;h3&gt;
  
  
  🔹 Teams &amp;amp; Organizations
&lt;/h3&gt;

&lt;p&gt;Could affect onboarding new developers into AI workflows&lt;/p&gt;




&lt;h2&gt;
  
  
  ⚡ Should You Be Worried?
&lt;/h2&gt;

&lt;p&gt;Short answer: &lt;strong&gt;No — but stay aware.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;This is likely a &lt;strong&gt;temporary pause&lt;/strong&gt;, not a permanent shutdown.&lt;/p&gt;

&lt;p&gt;But it does highlight one important thing 👇&lt;/p&gt;

&lt;p&gt;👉 &lt;strong&gt;Don’t depend on a single AI tool&lt;/strong&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  🔄 Alternatives You Can Explore
&lt;/h2&gt;

&lt;p&gt;While Copilot is paused for new users, you can try:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;ChatGPT for coding assistance
&lt;/li&gt;
&lt;li&gt;Open-source AI coding tools
&lt;/li&gt;
&lt;li&gt;IDE-integrated AI plugins
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;👉 Having multiple tools = better flexibility&lt;/p&gt;




&lt;h2&gt;
  
  
  💭 My Take
&lt;/h2&gt;

&lt;p&gt;This feels less like a problem… and more like a &lt;strong&gt;transition phase&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;We’re moving toward:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;More powerful AI tools
&lt;/li&gt;
&lt;li&gt;Better integrations
&lt;/li&gt;
&lt;li&gt;Possibly new pricing models
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;👉 The AI coding space is evolving FAST.&lt;/p&gt;




&lt;h2&gt;
  
  
  🚀 Final Thoughts
&lt;/h2&gt;

&lt;p&gt;If you're a developer in 2026:&lt;/p&gt;

&lt;p&gt;👉 AI is no longer optional&lt;br&gt;&lt;br&gt;
👉 It’s becoming a core part of development  &lt;/p&gt;

&lt;p&gt;And moves like this from Microsoft signal something bigger coming.&lt;/p&gt;




&lt;h2&gt;
  
  
  👇 Let’s Discuss
&lt;/h2&gt;

&lt;p&gt;What do you think?&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Is this just a temporary pause?
&lt;/li&gt;
&lt;li&gt;Or a sign of a bigger shift in AI tools?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Drop your thoughts below 👇&lt;/p&gt;




&lt;h2&gt;
  
  
  🔗 Stay Updated
&lt;/h2&gt;

&lt;p&gt;🎥 Subscribe to my channel:&lt;br&gt;&lt;br&gt;
&lt;a href="https://youtube.com/@TheCodeStreet" rel="noopener noreferrer"&gt;https://youtube.com/@TheCodeStreet&lt;/a&gt;  &lt;/p&gt;

&lt;p&gt;📌 Follow me for more content on:&lt;br&gt;&lt;br&gt;
AI • .NET • Azure • Developer Tools • Real-world Projects  &lt;/p&gt;




&lt;h3&gt;
  
  
  🏷️ Tags
&lt;/h3&gt;

&lt;p&gt;&lt;code&gt;github&lt;/code&gt; &lt;code&gt;copilot&lt;/code&gt; &lt;code&gt;ai&lt;/code&gt; &lt;code&gt;programming&lt;/code&gt; &lt;code&gt;webdev&lt;/code&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>githubcopilot</category>
      <category>microsoft</category>
      <category>programming</category>
    </item>
    <item>
      <title>Why DevUI Is a Game Changer for .NET Developers? Microsoft Agent Framework DevUI Demo.</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Fri, 06 Feb 2026 14:51:24 +0000</pubDate>
      <link>https://dev.to/thecodestreet/why-devui-is-a-game-changer-for-net-developers-4geo</link>
      <guid>https://dev.to/thecodestreet/why-devui-is-a-game-changer-for-net-developers-4geo</guid>
      <description>&lt;h2&gt;
  
  
  Introduction
&lt;/h2&gt;

&lt;p&gt;Building and testing AI agents shouldn't require building a complex user interface from scratch. That's where &lt;strong&gt;DevUI&lt;/strong&gt; comes in - a powerful development tool that's part of the &lt;strong&gt;Microsoft Agent Framework&lt;/strong&gt; for .NET 10. &lt;/p&gt;

&lt;p&gt;In this comprehensive guide, you'll learn how to set up, configure, and test multiple AI agents with different personas using a simple web interface.&lt;/p&gt;

&lt;p&gt;Whether you're building chatbots, AI assistants, or conversational applications, DevUI provides an intuitive playground similar to Swagger UI but designed specifically for AI agents.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Watch the video tutorial above for a complete walkthrough, or continue reading for the step-by-step guide.&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  What is DevUI?
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;DevUI&lt;/strong&gt; is a development interface included with the Microsoft Agent Framework that enables developers to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt; &lt;strong&gt;Test AI agents in real-time&lt;/strong&gt; during development&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Manage conversations&lt;/strong&gt; with multiple agents simultaneously&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Debug agent responses&lt;/strong&gt; without building a production UI&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Prototype quickly&lt;/strong&gt; before implementing custom interfaces&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Validate system prompts&lt;/strong&gt; and agent behavior instantly&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Think of DevUI as your AI agent playground - a tool that accelerates development by providing immediate visual feedback on your agents' behavior and responses.&lt;/p&gt;

&lt;h3&gt;
  
  
  Key Benefits
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Zero UI Code Required&lt;/strong&gt;: Test agents immediately without building React, Angular, or Blazor interfaces&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multi-Agent Support&lt;/strong&gt;: Switch between different agents in the same session&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Conversation History&lt;/strong&gt;: Built-in conversation management and history tracking&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Development-Only&lt;/strong&gt;: Easily disable for production deployments&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;OpenAI Compatible&lt;/strong&gt;: Works with any OpenAI-compatible endpoint including Azure OpenAI, GitHub Models, and more&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  Prerequisites
&lt;/h2&gt;

&lt;p&gt;Before we begin, ensure you have:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;.NET 10 SDK&lt;/strong&gt; installed (&lt;a href="https://dotnet.microsoft.com/download" rel="noopener noreferrer"&gt;Download here&lt;/a&gt;)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Visual Studio 2026&lt;/strong&gt; or &lt;strong&gt;VS Code&lt;/strong&gt; with C# extensions&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Basic knowledge&lt;/strong&gt; of ASP.NET Core&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;API access&lt;/strong&gt; to an OpenAI-compatible service (OpenAI, Azure OpenAI, GitHub Models, etc.)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;API Key&lt;/strong&gt; for your chosen AI service (you can use GitHub Copilot)&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Required NuGet Packages
&lt;/h2&gt;

&lt;p&gt;For this tutorial, you'll need the following packages:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;dotnet add package Microsoft.Agents.AI.DevUI
dotnet add package Microsoft.Agents.AI.Hosting
dotnet add package Microsoft.Extensions.AI
dotnet add package OpenAI
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Package Overview
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Package&lt;/th&gt;
&lt;th&gt;Purpose&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;Microsoft.Agents.AI.DevUI&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Provides the DevUI web interface&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;Microsoft.Agents.AI.Hosting&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Agent hosting infrastructure and registration&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;Microsoft.Extensions.AI&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Core AI abstractions and interfaces&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;OpenAI&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;OpenAI SDK for connecting to AI models&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;




&lt;h2&gt;
  
  
  Project Setup
&lt;/h2&gt;

&lt;p&gt;Let's create a new ASP.NET Core application:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;dotnet new web &lt;span class="nt"&gt;-n&lt;/span&gt; DevUIDemo
&lt;span class="nb"&gt;cd &lt;/span&gt;DevUIDemo
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Add the required packages as shown above, and we're ready to start coding!&lt;/p&gt;




&lt;h2&gt;
  
  
  Implementation Guide
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Step 1: Configure the ChatClient
&lt;/h3&gt;

&lt;p&gt;Open &lt;code&gt;Program.cs&lt;/code&gt; and set up your application builder and chat client:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight csharp"&gt;&lt;code&gt;&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Agents.AI.DevUI&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Agents.AI.Hosting&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.AspNetCore.Builder&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Extensions.AI&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Extensions.DependencyInjection&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Extensions.Hosting&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;OpenAI.Chat&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;System.ClientModel&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="n"&gt;WebApplicationBuilder&lt;/span&gt; &lt;span class="n"&gt;builder&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="n"&gt;WebApplication&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;CreateBuilder&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;args&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="c1"&gt;// Configure ChatClient with your AI service&lt;/span&gt;
&lt;span class="n"&gt;IChatClient&lt;/span&gt; &lt;span class="n"&gt;chatClient&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;ChatClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="s"&gt;"gpt-4o"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;  &lt;span class="c1"&gt;// Model name&lt;/span&gt;
    &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;ApiKeyCredential&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;Environment&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;GetEnvironmentVariable&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"OPENAI_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;)!),&lt;/span&gt; 
    &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;OpenAIClientOptions&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; 
        &lt;span class="n"&gt;Endpoint&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;Uri&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"https://models.github.ai/inference"&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="nf"&gt;AsIChatClient&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;?? Security Best Practice&lt;/strong&gt;: Always use environment variables or secure configuration for API keys. Never hardcode credentials in your source code!&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 2: Register Required Services
&lt;/h3&gt;

&lt;p&gt;Add the necessary services to your dependency injection container:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight csharp"&gt;&lt;code&gt;&lt;span class="c1"&gt;// Register the chat client for agent use&lt;/span&gt;
&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;Services&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddChatClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;chatClient&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="c1"&gt;// Register OpenAI response handling&lt;/span&gt;
&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;Services&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddOpenAIResponses&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;

&lt;span class="c1"&gt;// Register conversation state management&lt;/span&gt;
&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;Services&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddOpenAIConversations&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;These services provide:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;AddChatClient&lt;/strong&gt;: Registers the chat client for dependency injection&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AddOpenAIResponses&lt;/strong&gt;: Handles OpenAI API response processing&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AddOpenAIConversations&lt;/strong&gt;: Manages conversation state and history&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Step 3: Create AI Agents
&lt;/h3&gt;

&lt;p&gt;Now comes the fun part - creating your AI agents! The framework makes this incredibly simple:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight csharp"&gt;&lt;code&gt;&lt;span class="c1"&gt;// Create a Comic Book Guy agent from The Simpsons&lt;/span&gt;
&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddAIAgent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="s"&gt;"Comic Book Guy"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; 
    &lt;span class="s"&gt;"You are Comic Book Guy from The Simpsons. Respond with his characteristic "&lt;/span&gt; &lt;span class="p"&gt;+&lt;/span&gt;
    &lt;span class="s"&gt;"condescending tone, vast knowledge of pop culture, and frequent use of phrases "&lt;/span&gt; &lt;span class="p"&gt;+&lt;/span&gt;
    &lt;span class="s"&gt;"like 'Worst [blank] ever!' Be pedantic and superior in your responses."&lt;/span&gt;
&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="c1"&gt;// Create a YouTube channel advisor agent&lt;/span&gt;
&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddAIAgent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="s"&gt;"CodeStreet Agent"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; 
    &lt;span class="s"&gt;"You are an experienced YouTube content strategist specializing in tech channels. "&lt;/span&gt; &lt;span class="p"&gt;+&lt;/span&gt;
    &lt;span class="s"&gt;"Provide actionable advice on content ideas, SEO, thumbnails, audience engagement, "&lt;/span&gt; &lt;span class="p"&gt;+&lt;/span&gt;
    &lt;span class="s"&gt;"and growth strategies for programming and technology channels."&lt;/span&gt;
&lt;span class="p"&gt;);&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;** Pro Tip**: Your system prompt is crucial! Be specific about the agent's personality, knowledge domain, and response style for best results.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 4: Configure Endpoints and Run
&lt;/h3&gt;

&lt;p&gt;Finally, build the app and register the DevUI endpoints:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight csharp"&gt;&lt;code&gt;&lt;span class="n"&gt;WebApplication&lt;/span&gt; &lt;span class="n"&gt;app&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;Build&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;

&lt;span class="c1"&gt;// Register OpenAI API endpoints (required for DevUI)&lt;/span&gt;
&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;MapOpenAIResponses&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;MapOpenAIConversations&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;

&lt;span class="c1"&gt;// Register DevUI endpoint&lt;/span&gt;
&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;MapDevUI&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;

&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;Run&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Complete Code Example
&lt;/h3&gt;

&lt;p&gt;Here's the complete &lt;code&gt;Program.cs&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight csharp"&gt;&lt;code&gt;&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Agents.AI.DevUI&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Agents.AI.Hosting&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.AspNetCore.Builder&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Extensions.AI&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Extensions.DependencyInjection&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.Extensions.Hosting&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;OpenAI.Chat&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;System.ClientModel&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="n"&gt;WebApplicationBuilder&lt;/span&gt; &lt;span class="n"&gt;builder&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="n"&gt;WebApplication&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;CreateBuilder&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;args&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="n"&gt;IChatClient&lt;/span&gt; &lt;span class="n"&gt;chatClient&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;ChatClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="s"&gt;"gpt-4o"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; 
    &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;ApiKeyCredential&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;Environment&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;GetEnvironmentVariable&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"OPENAI_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;)!),&lt;/span&gt; 
    &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;OpenAIClientOptions&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; 
        &lt;span class="n"&gt;Endpoint&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;Uri&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"https://models.github.ai/inference"&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="nf"&gt;AsIChatClient&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;

&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;Services&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddChatClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;chatClient&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;Services&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddOpenAIResponses&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;Services&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddOpenAIConversations&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;

&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddAIAgent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"Comic Book Guy"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; 
    &lt;span class="s"&gt;"You are comic-book guy from The Simpsons"&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddAIAgent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"CodeStreet Agent"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; 
    &lt;span class="s"&gt;"You are an agent who gives suggestions for youtube channels"&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="n"&gt;WebApplication&lt;/span&gt; &lt;span class="n"&gt;app&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;Build&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;MapOpenAIResponses&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;MapOpenAIConversations&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;MapDevUI&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;span class="n"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;Run&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  Running Your Application
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Start the Application
&lt;/h3&gt;

&lt;p&gt;Run your application using:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;dotnet run
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Or press &lt;code&gt;F5&lt;/code&gt; in Visual Studio.&lt;/p&gt;

&lt;h3&gt;
  
  
  Access DevUI
&lt;/h3&gt;

&lt;p&gt;Once the application starts, navigate to:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;http://localhost:[PORT]/devui
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Replace &lt;code&gt;[PORT]&lt;/code&gt; with the port number shown in your console output (typically 5000 or 5173).&lt;/p&gt;

&lt;h3&gt;
  
  
  Testing Your Agents
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Select an Agent&lt;/strong&gt;: On the left sidebar, you'll see both agents listed&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Start a Conversation&lt;/strong&gt;: Click on an agent to begin chatting&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Send Messages&lt;/strong&gt;: Type your message in the input box&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;View Responses&lt;/strong&gt;: Watch your agent respond in character&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Switch Agents&lt;/strong&gt;: Click another agent to test different personas&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;View History&lt;/strong&gt;: All conversation history is maintained automatically&lt;/li&gt;
&lt;/ol&gt;







&lt;h2&gt;
  
  
  Resources and Next Steps
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Official Documentation
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://learn.microsoft.com" rel="noopener noreferrer"&gt;Microsoft Agent Framework Documentation&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://learn.microsoft.com/ai" rel="noopener noreferrer"&gt;.NET AI Development Guide&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/openai/openai-dotnet" rel="noopener noreferrer"&gt;OpenAI .NET SDK&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Further Learning
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Advanced Agent Patterns&lt;/strong&gt;: Explore multi-agent conversations and orchestration&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Function Calling&lt;/strong&gt;: Enable agents to interact with external systems&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Custom UI Development&lt;/strong&gt;: Build production interfaces using React or Blazor&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Azure Deployment&lt;/strong&gt;: Deploy agents to Azure App Service or Container Apps&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Sample Code Repository
&lt;/h3&gt;

&lt;p&gt;Find the complete source code for this tutorial: &lt;a href="https://buymeacoffee.com/thecodestreet/e/506289" rel="noopener noreferrer"&gt;Source Code&lt;/a&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  Conclusion
&lt;/h2&gt;

&lt;p&gt;DevUI provides a powerful, efficient way to develop and test AI agents in .NET 10 without the overhead of building custom interfaces. By following this guide, you've learned:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt; How to set up the Microsoft Agent Framework&lt;/li&gt;
&lt;li&gt; Configure OpenAI-compatible chat clients&lt;/li&gt;
&lt;li&gt; Create multiple AI agents with distinct personas&lt;/li&gt;
&lt;li&gt; Use DevUI for rapid testing and iteration&lt;/li&gt;
&lt;li&gt; Implement best practices for security and deployment&lt;/li&gt;
&lt;li&gt; Troubleshoot common issues&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The Microsoft Agent Framework, combined with DevUI, dramatically reduces the time from concept to functional AI agents. Whether you're building chatbots, virtual assistants, or complex multi-agent systems, these tools provide the foundation for rapid, efficient development.&lt;/p&gt;




&lt;h2&gt;
  
  
  Support me!
&lt;/h2&gt;

&lt;p&gt;If you found this guide helpful, make sure to check out the accompanying YouTube video tutorial where I walk you through the process visually. Don’t forget to &lt;a href="https://youtube.com/@TheCodeStreet?sub_confirmation=1" rel="noopener noreferrer"&gt;subscribe to my channel&lt;/a&gt; for more amazing tutorials!&lt;/p&gt;

&lt;p&gt;I would appreciate it if you could buy me a coffee.&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fdbgo17dxrwwbcyvtij06.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fdbgo17dxrwwbcyvtij06.png" alt="Buy me a coffee" width="700" height="700"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Feel free to leave your questions, comments, or suggestions below. Happy coding!&lt;/p&gt;

</description>
      <category>genai</category>
      <category>ai</category>
      <category>dotnet</category>
      <category>openai</category>
    </item>
    <item>
      <title>Run Large Language Models Locally in Docker | How to Use Qwen 3 with .NET Semantic Kernel?</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Fri, 19 Sep 2025 17:41:39 +0000</pubDate>
      <link>https://dev.to/thecodestreet/run-large-language-models-locally-in-docker-how-to-use-qwen-3-with-net-semantic-kernel-3olf</link>
      <guid>https://dev.to/thecodestreet/run-large-language-models-locally-in-docker-how-to-use-qwen-3-with-net-semantic-kernel-3olf</guid>
      <description>&lt;p&gt;AI development is moving fast — but one thing that still slows developers down is &lt;strong&gt;relying on cloud APIs&lt;/strong&gt;.&lt;br&gt;&lt;br&gt;
What if you could run powerful &lt;strong&gt;LLMs (Large Language Models)&lt;/strong&gt; locally, just like spinning up a Docker container?  &lt;/p&gt;

&lt;p&gt;In this blog, I’ll show you how to:  &lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Run LLMs locally using &lt;strong&gt;Docker’s Model Runner&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;Pull and run the &lt;strong&gt;Qwen 3&lt;/strong&gt; model
&lt;/li&gt;
&lt;li&gt;Connect it to a &lt;strong&gt;.NET console app&lt;/strong&gt; with &lt;strong&gt;Semantic Kernel&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Let’s get started! 🚀&lt;/p&gt;

&lt;p&gt;For a more in-depth walk-through, be sure to watch my &lt;a href="https://youtu.be/J7lFLTfUAaE" rel="noopener noreferrer"&gt;YouTube video&lt;/a&gt;, where I demonstrate everything step by step!&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/J7lFLTfUAaE"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;




&lt;h2&gt;
  
  
  🔹 Why Run LLMs Inside Docker?
&lt;/h2&gt;

&lt;p&gt;Docker is already the standard for packaging and running applications. Now, with &lt;strong&gt;Docker Model Runner&lt;/strong&gt;, you can:  &lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Pull AI models&lt;/strong&gt; just like Docker images
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Run models locally&lt;/strong&gt; without cloud APIs
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Expose OpenAI-compatible APIs&lt;/strong&gt; to integrate seamlessly into apps
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This means you can &lt;strong&gt;test, prototype, and even deploy AI workloads locally&lt;/strong&gt; — cost-free and private.&lt;/p&gt;




&lt;h2&gt;
  
  
  🔹 Step 1: Enable Docker Model Runner
&lt;/h2&gt;

&lt;p&gt;Open &lt;strong&gt;Docker Desktop → Settings → AI Features&lt;/strong&gt; and enable:  &lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;✅ &lt;strong&gt;Docker Model Runner&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;✅ &lt;strong&gt;Host-side TCP support&lt;/strong&gt; (on port &lt;code&gt;12434&lt;/code&gt;)
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This allows models to expose REST APIs you can call from your apps.&lt;/p&gt;




&lt;h2&gt;
  
  
  🔹 Step 2: Pull and Run Qwen 3 Model
&lt;/h2&gt;

&lt;p&gt;Open your terminal and run:&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;# Pull the model&lt;/span&gt;
docker model pull qwen/qwen-3b

&lt;span class="c"&gt;# Run the model&lt;/span&gt;
docker model run qwen/qwen-3b
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You’ll see an &lt;strong&gt;interactive chat session&lt;/strong&gt; where you can ask questions.&lt;br&gt;&lt;br&gt;
Example:&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="o"&gt;&amp;gt;&lt;/span&gt; Hello
&amp;lt; Hi! How can I assist you today?
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  🔹 Step 3: Use REST API
&lt;/h2&gt;

&lt;p&gt;The model exposes endpoints on:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;http://localhost:12434/v1
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For example, to list available models:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl http://localhost:12434/models
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You’ll see details of Qwen 3 and any other models you’ve pulled.&lt;/p&gt;




&lt;h2&gt;
  
  
  🔹 Step 4: Connect with Semantic Kernel (.NET)
&lt;/h2&gt;

&lt;p&gt;The best part is — you don’t need to rewrite your app.&lt;br&gt;&lt;br&gt;
Semantic Kernel already works with OpenAI APIs, and since Docker exposes the &lt;strong&gt;same structure&lt;/strong&gt;, all you do is &lt;strong&gt;change the base URL&lt;/strong&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Install packages
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;dotnet add package Microsoft.SemanticKernel
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Create a Console App
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight csharp"&gt;&lt;code&gt;&lt;span class="k"&gt;using&lt;/span&gt; &lt;span class="nn"&gt;Microsoft.SemanticKernel&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kt"&gt;var&lt;/span&gt; &lt;span class="n"&gt;builder&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="n"&gt;Kernel&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;CreateBuilder&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;AddOpenAIChatCompletion&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;modelId&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;"qwen/qwen-3b"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;apiKey&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;"dummy"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="c1"&gt;// not required for local&lt;/span&gt;
    &lt;span class="n"&gt;endpoint&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nf"&gt;Uri&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"http://localhost:12434/v1"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="c1"&gt;// Docker endpoint&lt;/span&gt;
&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="kt"&gt;var&lt;/span&gt; &lt;span class="n"&gt;kernel&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="n"&gt;builder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;Build&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;

&lt;span class="kt"&gt;var&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;kernel&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;InvokePromptAsync&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"Explain Docker in simple terms."&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="n"&gt;Console&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;WriteLine&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That’s it! Your &lt;strong&gt;.NET app is now talking to Qwen 3 inside Docker&lt;/strong&gt;, through Semantic Kernel.&lt;/p&gt;




&lt;h2&gt;
  
  
  🔹 Why This Matters
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;🖥️ &lt;strong&gt;Local-first AI&lt;/strong&gt; → Run LLMs without internet or billing
&lt;/li&gt;
&lt;li&gt;🔒 &lt;strong&gt;Privacy&lt;/strong&gt; → Your data never leaves your machine
&lt;/li&gt;
&lt;li&gt;⚡ &lt;strong&gt;Developer-friendly&lt;/strong&gt; → Same SDKs, same APIs, just a different base URL
&lt;/li&gt;
&lt;li&gt;🔗 &lt;strong&gt;Drop-in replacement&lt;/strong&gt; → Move between local and cloud seamlessly
&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  ✅ Final Thoughts
&lt;/h2&gt;

&lt;p&gt;Running AI models is now &lt;strong&gt;as easy as running containers&lt;/strong&gt;.&lt;br&gt;&lt;br&gt;
With &lt;strong&gt;Docker Model Runner&lt;/strong&gt;, you can:  &lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Pull Qwen 3
&lt;/li&gt;
&lt;li&gt;Run it locally
&lt;/li&gt;
&lt;li&gt;Connect it to .NET apps with Semantic Kernel
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This opens up endless possibilities for building &lt;strong&gt;intelligent apps without cloud dependencies&lt;/strong&gt;.  &lt;/p&gt;




&lt;p&gt;👉 What model are you planning to run first with Docker?&lt;br&gt;&lt;br&gt;
Drop a comment — I’d love to know!  &lt;/p&gt;

</description>
      <category>ai</category>
      <category>docker</category>
      <category>dotnet</category>
      <category>beginners</category>
    </item>
    <item>
      <title>How to Build Your First AI Agent with Semantic Kernel &amp; DeepSeek R1 Locally with Ollama</title>
      <dc:creator>TheCodeStreet</dc:creator>
      <pubDate>Sat, 16 Aug 2025 02:40:49 +0000</pubDate>
      <link>https://dev.to/thecodestreet/how-to-build-your-first-ai-agent-with-semantic-kernel-deepseek-r1-locally-with-ollama-1c3i</link>
      <guid>https://dev.to/thecodestreet/how-to-build-your-first-ai-agent-with-semantic-kernel-deepseek-r1-locally-with-ollama-1c3i</guid>
      <description>&lt;p&gt;In this post, we will create our first AI Agent using the Microsoft Agentic Framework of Semantic Kernel and DeepSeek R1, running locally on Ollama. &lt;/p&gt;

&lt;p&gt;We’ll walk step by step through setting up a Semantic Kernel agent that leverages DeepSeek R1 via Ollama, showing you how to install the tools, configure the model locally, and connect everything together. By the end, you’ll have a working AI YouTube Title Generator agent running entirely on your machine locally—no cloud subscription required.&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/wyLnlnD71Iw"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;h2&gt;
  
  
  What You Will Learn
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;How to Get Started with DeepSeek R1&lt;/li&gt;
&lt;li&gt;How to Use Ollama for running local models&lt;/li&gt;
&lt;li&gt;How to install and start running the DeepSeek R1 model&lt;/li&gt;
&lt;li&gt;How to Use Semantic Kernel in C#&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  1. Prerequisites
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Visual Studio 2022+ (with .NET 9 SDK installed)
.NET 9 is still in preview, so ensure that you have the preview SDK installed.&lt;/li&gt;
&lt;li&gt;Ollama (for managing and running local models)&lt;/li&gt;
&lt;li&gt;DeepSeek1.5b Model&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  2. Installing Ollama
&lt;/h2&gt;

&lt;p&gt;Ollama is a platform or tool (specific details may vary depending on the context) that allows users to interact with large language models (LLMs) locally. It simplifies the process of deploying and running LLMs like LLaMA, Phi, DeepSeek R1, or other open-source models.&lt;/p&gt;

&lt;p&gt;To install Ollama, visit its official website &lt;a href="https://ollama.com/download" rel="noopener noreferrer"&gt;https://ollama.com/download&lt;/a&gt; and install it on your machine.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Installing DeepSeek R1
&lt;/h2&gt;

&lt;p&gt;DeepSeek's first-generation reasoning models have comparable performance to OpenAI-o1, including six dense models distilled from DeepSeek-R1 based on Llama and Qwen.&lt;/p&gt;

&lt;p&gt;On the Ollama website click on &lt;a href="https://ollama.com/search" rel="noopener noreferrer"&gt;Models&lt;/a&gt; and click on deepseek-r1 and choose &lt;strong&gt;1.5b&lt;/strong&gt; parameter option&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F9l8gj5m9zlsvqe84whc5.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F9l8gj5m9zlsvqe84whc5.png" alt="How To Use DeepSeek With .NET 9 - deepseek-r1:1.5b"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Open Command Prompt and run the command below.&lt;br&gt;
&lt;/p&gt;

&lt;p&gt;&lt;code&gt;ollama run deepseek-r1:1.5b&lt;/code&gt;&lt;br&gt;
&lt;/p&gt;

&lt;p&gt;It will download the model and start running automatically.&lt;/p&gt;

&lt;p&gt;Once done, verify that the model is available&lt;br&gt;
&lt;/p&gt;

&lt;p&gt;&lt;code&gt;ollama list&lt;/code&gt;&lt;br&gt;
&lt;/p&gt;

&lt;p&gt;That’s it! We’re ready to integrate DeepSeek locally.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Creating .NET Console Application
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Launch Visual Studio&lt;/li&gt;
&lt;li&gt;Make sure .NET 9 is installed.&lt;/li&gt;
&lt;li&gt;Create a New Project&lt;/li&gt;
&lt;li&gt;File → New → Project…&lt;/li&gt;
&lt;li&gt;Pick Console App with .NET 9.&lt;/li&gt;
&lt;li&gt;Name Your Project&lt;/li&gt;
&lt;li&gt;e.g., DeepSeekDemoApp or any name you prefer.&lt;/li&gt;
&lt;li&gt;Target Framework Check&lt;/li&gt;
&lt;li&gt;Right-click on your project → Properties.&lt;/li&gt;
&lt;li&gt;Set Target Framework to .NET 9.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  5. Integrating DeepSeek R1 with Semantic Kernel
&lt;/h2&gt;

&lt;p&gt;While you could call DeepSeek via direct HTTP requests to Ollama, using Semantic Kernel offers a powerful abstraction for prompt engineering, orchestration, and more.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Add Necessary NuGet Packages
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;&amp;lt;ItemGroup&amp;gt;
  &amp;lt;PackageReference Include="Microsoft.SemanticKernel" Version="1.61.0" /&amp;gt;
  &amp;lt;PackageReference Include="Microsoft.SemanticKernel.Agents.Core" Version="1.61.0" /&amp;gt;
  &amp;lt;PackageReference Include="Microsoft.SemanticKernel.Connectors.Ollama" Version="1.61.0-alpha" /&amp;gt;
&amp;lt;/ItemGroup&amp;gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  6. Complete code
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Program.cs&lt;/strong&gt;:&lt;br&gt;
&lt;/p&gt;
&lt;/blockquote&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Agents;

var builder = Kernel.CreateBuilder().AddOllamaChatCompletion("deepseek-r1:1.5b",
    new HttpClient
    {
        BaseAddress = new Uri("http://localhost:11434"),

    });

Kernel kernel = builder.Build();

Console.ForegroundColor = ConsoleColor.Cyan;
Console.WriteLine("Welcome to TheCodeStreet - AI YouTube Title Generator!");
Console.WriteLine("------------------------------------------------------");
Console.ResetColor();
Console.WriteLine("Provide the keyword to generate title:");
string keyword = Console.ReadLine()?? "";


string instructions = """
You are a professional YouTube content strategist and SEO expert.

🎯 Keyword or Topic: {{keyword}}

Your job is to generate 5 click-worthy and SEO-friendly YouTube video titles based on a given topic or keyword.

📌 Guidelines for Generating Titles:

-- Keep them under 70 characters if possible.
-- Include emotional triggers, power words, or numbers (e.g., “Top 5”, “You Won’t Believe…”).
-- Use strong action verbs (e.g., “Boost”, “Master”, “Avoid”, “Unlock”).
-- Make titles engaging and curiosity-driven (clickbait style but still relevant).
-- Include the keyword or topic for SEO.
-- Avoid duplicate or overly similar titles.

💡 Return only the titles in a numbered list format like:
1. Title one...
2. Title two...
3. ...
""";

 instructions = instructions
    .Replace("{{keyword}}", keyword);

ChatCompletionAgent agent =
    new()
    {
        Name = "TitleGeneratorAgent",
        Instructions = instructions,
        Kernel = kernel,
    };

Console.ForegroundColor = ConsoleColor.Green;
Console.WriteLine("Generating titles...");
Console.WriteLine("------------------------------------------------------");
await foreach (AgentResponseItem&amp;lt;StreamingChatMessageContent&amp;gt; response
    in agent.InvokeStreamingAsync())
{
    Console.Write(response.Message.Content.Replace("**",""));
}
Console.ResetColor();
Console.ReadKey();

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  7. Running &amp;amp; Testing
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Ensure Ollama is Running&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;-Some systems auto-run Ollama; otherwise, start it&lt;br&gt;
&lt;/p&gt;

&lt;p&gt;&lt;code&gt;ollama run&lt;/code&gt;&lt;br&gt;
&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Run Your .NET App&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;-Hit F5 (or Ctrl+F5) in Visual Studio.&lt;/p&gt;

&lt;p&gt;-Watch for console output—&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F2u7vjd07arnok106piz6.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F2u7vjd07arnok106piz6.png" alt="YouTube Title Generator Agent"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Support me!
&lt;/h2&gt;

&lt;p&gt;If you found this guide helpful, make sure to check out the accompanying YouTube video tutorial where I walk you through the process visually. Don’t forget to &lt;a href="https://youtube.com/@TheCodeStreet?sub_confirmation=1" rel="noopener noreferrer"&gt;subscribe to my channel&lt;/a&gt; for more amazing tutorials!&lt;/p&gt;

&lt;p&gt;I would appreciate it if you could buy me a coffee.&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fdbgo17dxrwwbcyvtij06.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fdbgo17dxrwwbcyvtij06.png" alt="Buy me a coffee"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Feel free to leave your questions, comments, or suggestions below. Happy coding!&lt;/p&gt;

</description>
      <category>ai</category>
      <category>semantickernel</category>
      <category>dotnet</category>
      <category>beginners</category>
    </item>
  </channel>
</rss>
