Cindy, a free open-source agent, lets you swap Claude Code and Codex mid-task
Cindy is an Apache-2.0 desktop and mobile app that runs two coding agents under one workflow. It is early and untested by outsiders, so read the code first.
The short version
- Cindy runs locally and works with your real files and logged-in apps.
- Claude Code and Codex are the first two supported harnesses; more are planned, plus a native one.
- You can use the official Cindy service, a plan you already pay for, your own API keys or local models.
- The client is under Apache-2.0; the backend lives in a separate repository.
- Usage analytics are on by default but can be switched off in settings.
Why it matters
Coding agents are improving quickly, and many people now pay for more than one. A layer that keeps your setup steady while the agent underneath changes could save rework. But a tool that launches other agents with access to your machine deserves a careful read before it gets a real project.
One agent, several engines
Cindy is an open-source AI agent that runs on your own computer and works with your actual files and the apps you are already logged into. As github.com reported in the project's own description, its central idea is to bring several models, tools and "harnesses" together in one place.
A harness is the wrapper around an AI model that hands it tools and runs the loop of acting, checking and trying again. Claude Code (from Anthropic) and Codex (from OpenAI) are two such coding agents. They are the first two harnesses Cindy supports. The project says more are coming, and that it is working on a native harness of its own.
What swapping actually means
The pitch is that you can change the model or the harness partway through a task. Your workspace, memory, skills and tools are meant to carry over unchanged. In practice, that would let you compare two agents on the same job without rebuilding your setup each time.
The project goes further. It says a single task can be planned, run in parallel and reviewed by agents using different combinations of harness and model. Cindy can also drive your browser, computer and phone, and it can accept work from messaging apps and from schedules.
All of this comes from the project's own description. We have not tested it, and the source offers no benchmarks or outside reviews.
Paying for it, and what is open
The client is free, and its source code is open under the Apache-2.0 license. You have four ways to supply the AI itself:
- Sign in to the official Cindy service, where usage is deducted.
- Authorize the Claude Code or Codex plan you already pay for, so there is no second bill.
- Connect your own API keys.
- Use local models.
The project says "open" should mean changeable, not just visible. It lists memory (correct her once and she remembers, across harnesses), skills you teach once and reuse, scheduled automation, and MCP connections to your internal tools. MCP, the Model Context Protocol, is a standard way to plug outside services into an agent. Plugins and team sharing of skills are described as still being built.
One boundary to know about: this repository holds only the desktop and mobile client. The backend service is in a separate repository. By default, the client connects to Cindy's official cloud services. A "Skip Sign-In" option runs local agents without an account, but server-backed features are unavailable in that mode.
Read the privacy fine print
The project is unusually detailed about analytics, which helps. Official builds include a usage-statistics tool called TapDB, which records device identifiers, operating system and app version, and startup and activity events. After sign-in these events are tied to your account ID, so it is not fully anonymous. The project says it does not collect chat content, file content or working-directory data.
The statistics setting is on by default. On desktop you can turn it off under Settings, then About, then Usage statistics. Mobile has an equivalent switch. The README also notes a known quirk: on mobile, a restored enterprise single-sign-on login can mark even a new installation as consented. Separate heartbeats and diagnostic logs are not controlled by that switch. Desktop heartbeats are sent only during a verified cloud sign-in, and crash dumps stay local.
Early days, so go carefully
Cindy had only 2 points on Hacker News and no comments when it was listed, so there is almost no outside feedback yet. Our hosts' read is that it is worth a look for builders who already use both Claude Code and Codex.
The caution is the same one raised in today's lead story about agents with terminal access. An agent can install packages, run scripts and connect to servers. A tool that launches such agents is another layer with that reach, so read its code and try it in a sandbox, not on your real projects.
For contributors, the project asks for Node.js 22.x, pnpm 10.x and Git LFS. Changes go through pull requests with a sign-off on each commit, and no contributor license agreement is required.
What to watch
- Will independent users test whether swapping between Claude Code and Codex keeps memory and context intact?
- Which additional harnesses will be added, and what will the native one look like?
- Will the plugin marketplace and team skill sharing, both still in progress, actually ship?
Try it
Before installing, open the repository, read the analytics section, and note the Settings, About, Usage statistics switch. If you try Cindy, use a throwaway folder, not your real projects.
Sources and links
- Reportinggithub.com: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by github.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
