Dependencies, priorities and continuous developer mode all work — a task unblocks when the one it depends on is approved, and the developer moves straight to the next item. What does not work autonomously yet is the loop itself: agents cannot schedule their own checks, so someone runs the status commands on the 120-second interval, and the maximum iteration count is 500. Automation scripts ship with the repository but need setup. Treat it as a project that is honest about being in progress.
An MCP server for AI-to-AI collaboration, designed so Claude and Gemini can work a project between them without a human relaying messages. State lives in JSON files on disk — tasks, missions, project state, plans, loop state and tickets — so both agents read the same board across sessions.
- A single entry point that loads existing context, resumes or creates the project plan, surfaces blockers and starts the work loop — `init`
- Directive creation from the CTO side, one at a time or batched, carrying priority and `dependsOn` relationships so tasks unblock automatically when their predecessors are approved — `send_directive`, `send_batch_directives`
- The developer side: a priority-sorted view of assigned work, submission of completed tasks, and asynchronous questions when a spec is unclear — `get_all_tasks`, `submit_work`, `ask_question`
- Review and plan progression from the CTO side — `review_work`, `create_project_plan`, `update_plan_progress`
- Loop status for each agent, which is how continuous work mode is driven — `get_loop_status`
- A task status model of available, blocked, in_progress, in_review and completed, plus a ticketing system for bugs, enhancements and tech debt
No account and no key. Clone the repository, run `npm install`, make `src/index.js` executable, then register it — a `.mcp.json` entry for Claude Code, or `~/.gemini/settings.json` for Gemini. The package is `ai-collab-mcp` (1.0.0 in package.json).
Build from source — clone the repository and build it, then point your client at the binary