Do not reveal secrets
The agent may paste from local clipboard or guide the operator, but it must not print unlock codes or tokens into chat.
Agent + Computer Use Tutorial
This guide is for advanced users who want a local coding agent to help configure the ChatGPT connector UI. The human operator still owns all secrets and final authorization.
Operator rules for agents
Stop at safety interstitials
If the browser asks the human to trust a tunnel or site, the agent should pause and ask the human to click.
Verify after setup
The agent should test `/health`, OAuth metadata, and a minimal `file.list` or `file.read_path` call after authorization.
Read policy first
After linking, call `policy.read` before broad file operations and confirm skill roots use `~/.codex/skills`.
Route local skills
Use `skill.route` or `skill.search` before reading a local `SKILL.md`, then use `skill.bundle` for references.
Suggested agent prompt
You are configuring ChatGPT2LocalBridge with Computer Use.
1. Verify local health at http://127.0.0.1:3838/health.
2. Verify OAuth metadata at the public tunnel URL.
3. Open ChatGPT connector settings.
4. Create a custom connector named ChatGPT2LocalBridge.
5. Use URL https://YOUR-DOMAIN/mcp and OAuth for public tunnels.
6. If the OAuth page asks for an unlock code, ask the human operator to provide or approve it.
7. Never print tokens or unlock codes.
8. After linking, send a minimal file.list or file.read_path test inside the approved workspace.
9. Call policy.read and verify allowedProjectRoots and skillRoots.
10. Use skill.route before reading local skills.
11. For Linux paths, create a separate Linux connector instead of reusing the Mac connector.