Leading the Claude and Codex desktop apps phase by phase, instead of the CLI

When people automate coding agents, they usually script the CLI. I built OpenDaemon to lead the agent inside the desktop apps instead (the Claude desktop app, the Codex desktop app and the Antigravity IDE), because that's where most people already work, and because watching the agent on screen is how you learn to trust it.

The loop

1. Plan first. You describe the app once. A planner writes phases plus a checklist: what you asked for, plus what anyone would expect from that kind of thing. No code is written before the plan exists.
2. Brief each phase. For every phase, OpenDaemon writes a full brief and sends it to a fresh conversation in your coding app.
3. Check, don't trust. It reads the report and screenshots. A phase that doesn't open or run goes back to be fixed, never marked done.
4. Playtest last. The final phase runs what was built, walks the checklist item by item and saves screenshots. Anything it couldn't check is listed for you.
5. You stay in charge. Approve each phase, or approve the plan once with Full trust.

What I learned

• "Done" is the hard problem. Agents report success eagerly. Requiring something you can open at the end of every phase removed most false "done"s.
• Questions should come back, not be guessed. When the agent is unsure, the question comes back to you instead of being invented.
• Every project gets its own folder, and "Continue conversation" picks up where it stopped.

Try it

Free on the Microsoft Store (Windows 10/11): https://apps.microsoft.com/detail/9NRBTQJJXF4G?cid=devto

It runs on NVIDIA's free models or your own key. The AI Developer needs Chrome plus the Claude desktop app, the Codex desktop app or Antigravity. Feedback welcome, especially on how you decide an agent's work is really finished.