OpenChamber is an open source development environment for AI agents that work on code. Released on GitHub and reaching 132 points on Hacker News with 72 comments, the tool runs locally on the OpenCode SDK and keeps each session isolated.

For those bringing agents into daily work, this tool answers a need that has become concrete. As we covered on August 5 with session isolation in Claude Code, agentic governance has moved from theory to practice: controlling what an agent does, in which folder, and with which permissions, matters more than the model you put underneath it. OpenChamber brings this control to the moment when it really matters, while you develop and test, before sending the agent to production.

The distinguishing features: give the agent a final objective and it continues working on it turn after turn, even with the app closed. You can launch the same task across up to five models and keep the best result, or merge the strongest parts. You can point to an element of the running app and send the agent all the context behind that element. All from desktop, browser, phone, or VS Code, without losing the thread of the session.

The tool is free and open source. Code and sessions stay on your machine; remote browser access is protected with password and rotatable tunnels.

If you want to try it, download the app from openchamber.dev and follow the documentation to install the OpenCode SDK.

In detail

Coding agents have moved beyond demo stage. Claude Code, OpenAI’s Codex, Cursor: each has its own harness, interface, security rules. What was missing was an environment built around the idea that an agent works on multiple turns, on long objectives, and that every session should be isolated and observable before you trust it.

OpenChamber positions itself exactly there. It’s a development environment that acts as interface and runtime for agents based on the OpenCode SDK, an open source harness chosen by the team for the quality of the agentic experience.

Session isolation. Each agent works in its own space. If you launch five in parallel across five different models for the same task (the multi-run and fusion feature), they don’t interfere with each other. The diff they produce gets grouped into ordered steps, so you can read how the result was reached, not just the final result. This solves a real problem: when an agent makes a large change, understanding what it touched and why is often harder than writing it yourself.

Session goals. You set a target and the agent works on it until it reaches it, even if you close the app. It’s the difference between an assistant waiting for your next message and a worker with an objective. The limitation is obvious: the quality of the result depends on how well you defined the target, and an ambiguous objective produces ambiguous work.

Access everywhere, with protection. OpenChamber runs as a desktop app (macOS, Windows, Linux), as a PWA in the browser, and as a mobile app in beta. For remote access it uses a private relay with pairing via QR code, without opening ports on your router. Browser access is protected with password, and tunnels can be revoked. Privacy is declared: no code or session data is collected or sent elsewhere. Being open source, the privacy model is in the code.

The limitations. The tool is young. Testimonies from early users on Hacker News describe an already solid product, but direct comparison with mature environments is lacking. GitHub integration (issues, PRs, failed checks) is promised and partly implemented, but complete documentation is still in progress. The scheduled work feature (cron-based prompts) is interesting but needs to be verified on real cases before relying on it for production work. The mobile app is in beta. And the fact that it runs on the OpenCode SDK means you’re bound to that harness: if you use a different framework, you need to switch to this one.

What changes for you. If you’re using Claude Code or Codex and wondering how to test an agent before sending it to production, OpenChamber gives you an environment with isolation, multi-run, and diff observability. If you don’t yet have a coding agent in your workflow, this tool lowers the barrier because it brings together in a single interface what would otherwise require three or four separate tools. The signal running through the whole week of multiple agents is the same: the single tool matters less than the environment it runs in.