Open Source · AI Agent · October 2026

Cindy: The Open-Source AI Agent That Switches Harnesses Mid-Task

Cindy (makecindy/cindy) is an open-source AI agent that refuses to make you pick a side. Instead of locking one task to one tool, it lets you mix Claude Code, Codex, and Pi — switching harnesses or models mid-task while your workspace, memory, skills, and tools stay continuous. The pitch, in the project's own words: “Think it. Consider it done.” It trended on GitHub in late October 2026, and the one idea worth understanding is the harness layer.

Last updated: October 25, 2026. Star count and project status checked against github.com/makecindy/cindy and cindy.app on that date.

At a Glance

The Idea That Matters: The Harness Layer

A model produces tokens. Getting useful work out of one takes a lot of machinery around it: something that holds context across many turns, decides which tools to expose, runs them in a sandbox, streams progress, catches failures, and pauses to ask a human before anything irreversible. That surrounding execution system is the harness. Claude Code is one harness. Codex is another. Each has blind spots — good at some things, stubborn about others — and until now you could not just swap in a different one when you hit a wall without losing the thread.

Cindy's answer is to hold the task state above the harness boundary. Cindy keeps the workspace, the memory, the skills, the tools, and the acceptance criteria; the harness underneath becomes replaceable. One task can be planned, executed in parallel, and reviewed by agents on different harness-and-model combinations, and everything switches mid-task without losing context. If that layer works, the question stops being “which agent do I use?” and becomes “which combination fits this step of the task?”

One Task, Three Roles: The Showcase Crew

Cindy's own example makes the idea concrete. A single task is split into three roles, each on a different combination: Claude Code plans first (architecture, work breakdown, acceptance criteria); three workers execute in parallel — one on Codex, one on Claude Code, one on Codex again, each with a different model — covering implementation, regression tests, and docs; then Codex performs an independent review as a quality gate (diff review, edge cases, pass or rework). Harnesses and models pass the baton while Cindy holds the task together. Treat this as the project's demonstration, not an independently measured benchmark — the honest test is running your own task through it.

Local-First, and More Than Code

Cindy runs locally on your own machine, using your real files and logged-in apps — the local-first premise is part of the appeal for developers who do not want their work routed through someone else's cloud by default. And the ambition goes past code: the project says Cindy can drive your browser, computer, and phone, and take work from instant messages and schedules. That makes it closer to a general-purpose agent that does real work on your computer than to a coding assistant with a chat box.

For model access, Cindy avoids the double-billing trap: if you already pay for a Claude Code or Codex coding plan, you authorize that plan inside Cindy and keep using it there. Otherwise you can sign in to the official Cindy service, connect your own API keys, or use local models. The client itself is free. Service pricing lives on cindy.app and is not restated here, so check the official page before committing to the hosted route.

The Honest Caveats

Three things to keep straight. First, only the client is open source: the backend service lives in a separate repository and is not part of the monorepo, so “open-source agent” describes the apps and shared packages, not the whole hosted stack. Second, the Cindy-native harness is in development — a roadmap item, not something you can use today. Third, this is a young, fast-moving project: it started in July 2026, and its issue tracker is busy. That is normal for a trending open-source agent, but it means you should read the README on github.com/makecindy/cindy before installing and expect rough edges rather than a polished product.

If you follow the proactive-agent wave, Cindy sits next to Anthropic's Claude Duet — but where Duet is a hosted, approval-gated product for professionals, Cindy is the open, local-first counterpart for developers who want to hold the harness layer themselves. For the model side of the same stack, see our Voxtral 3 guide. And if you want to turn agent output into video, start with HAILUOAI text-to-video.

FAQ