<?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: CodeGo AI</title>
    <description>The latest articles on DEV Community by CodeGo AI (@codego-ai).</description>
    <link>https://dev.to/codego-ai</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%2F4258331%2F207fd783-7e86-429b-9acb-36fc06b9e9d9.png</url>
      <title>DEV Community: CodeGo AI</title>
      <link>https://dev.to/codego-ai</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/codego-ai"/>
    <language>en</language>
    <item>
      <title>I got tired of babysitting five coding agents, so I built a supervisor for them</title>
      <dc:creator>CodeGo AI</dc:creator>
      <pubDate>Sun, 22 Nov 2026 06:12:14 +0000</pubDate>
      <link>https://dev.to/codego-ai/i-got-tired-of-babysitting-five-coding-agents-so-i-built-a-supervisor-for-them-ko4</link>
      <guid>https://dev.to/codego-ai/i-got-tired-of-babysitting-five-coding-agents-so-i-built-a-supervisor-for-them-ko4</guid>
      <description>&lt;p&gt;I don't use one coding agent anymore. Depending on the task, I switch between Claude Code, Codex, Gemini CLI, OpenCode and a few others.&lt;/p&gt;

&lt;p&gt;The tools themselves are great. The problem is everything &lt;em&gt;around&lt;/em&gt; them: separate sessions, separate quotas, zero shared context. And the moment one of them hits a 429 mid-task, I'm the one copy-pasting context into another tool and re-explaining the problem from scratch. I was spending real time being a human message bus between my own tools.&lt;/p&gt;

&lt;p&gt;So I built &lt;strong&gt;Kooyu&lt;/strong&gt;: a local-first personal assistant that sits on top of 10+ coding harnesses — Claude Code, Codex, DeepSeek, Gemini CLI, OpenCode, Cursor, ZCode, Kimi Code, Grok Build, pi — and supervises them.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the supervisor actually does
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Routing.&lt;/strong&gt; Each task goes to the engine that fits it best, instead of whichever tab happens to be open.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Rate-limit handoff.&lt;/strong&gt; When an engine hits a 429, Kooyu packages the live context — what was tried, what changed, what's left — and hands the task to another engine. The work continues instead of dying.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Arena mode.&lt;/strong&gt; This is the feature I use most. Give Kooyu a problem and it races 2–3 engines on it in parallel, each in its own isolated Git worktree. You get a scorecard — diff plus test results — and merge the winner. It's a surprisingly good way to learn which engine is actually better at which kind of task, with evidence instead of vibes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The boring-but-important parts.&lt;/strong&gt; Persistent local memory across sessions, background watches that cost zero tokens, and decision cards you can approve from your phone or IM when an agent needs a yes/no.&lt;/p&gt;

&lt;h2&gt;
  
  
  Local-first is the point
&lt;/h2&gt;

&lt;p&gt;Everything runs on your machine. Code, memory and API keys never leave it. Sync between desktop, web, iOS, Android and IM is end-to-end encrypted (zero-knowledge), which is what makes "approve from your phone" safe to offer at all.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;Kooyu is MIT-licensed and open source. v1.0.0 just shipped.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx kooyu@latest
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Site: &lt;a href="https://kooyu.app" rel="noopener noreferrer"&gt;https://kooyu.app&lt;/a&gt;&lt;br&gt;
GitHub: &lt;a href="https://github.com/sh2001sh/Kooyu" rel="noopener noreferrer"&gt;https://github.com/sh2001sh/Kooyu&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;If you also juggle multiple coding agents, I'd love to hear it: what would you want a supervisor to take off your plate — and what would it have to prove before you'd trust it with a real repo?&lt;/p&gt;

</description>
      <category>productivity</category>
      <category>ai</category>
      <category>opensource</category>
      <category>devtools</category>
    </item>
  </channel>
</rss>
