chutingyean/Agent-Cockpit
WHITEBOARD WITH MCP FOR RESEARCH AI AGENTS
Local whiteboard and MCP server where every agent claim traces back to the code and run that produced it. A deterministic engine tracks evidence chains and flags stale claims when anything changes.
// KEY FEATURES
- Evidence chains down to JSON fields with script and dependency hashes
- Works with any MCP agent: Claude Code, Codex, Cursor, Gemini CLI
- Built-in code graph and rule-based review without extra installs
- Replay mode: append-only events let you scrub the board to any moment
// DEEP DIVE: HOW IT WORKS
Agent Cockpit consists of a local web board (port 4180) and an MCP server through which agents write plans, tasks, runs, and claims. Each run records the hash of its script, every reachable function, and dependent files; when any of those changes, the run goes stale and every claim built on it is flagged for recheck. All writes pass a single validator: non-compliant writes are rejected with a reason the agent can act on. Unreported file changes are detected automatically and block git commit until reported.
Unlike typical experiment trackers, verification is deterministic and LLM-free: rules are pure algorithms, zero tokens spent. A lightweight call graph (graphify-style) supplies dependency closures, communities, core nodes, and impact analysis; if graphify or open-code-review are installed they are used automatically. Agents request human decisions via cockpit_decide and wait; when evidence breaks, the cockpit raises a decision itself and writes a one-page brief via answer-me-with-html.
Requirements: Node.js 22.13+ (built-in node:sqlite), git, python3 for the demo. The board binds to 127.0.0.1, no external keys needed. Limitation: local-only, no built-in multi-machine sync; MCP config is generated by cockpit init for Claude Code, Cursor, VS Code, Gemini CLI, and the command for Codex is printed.
// USE CASES
Best for: Researchers and engineers running long-lived experiments with AI agents
- Run reproducible research sessions with multiple AI agents
- Track which conclusions became stale after code or data edits
- Rewind experiment history to any step for audit
// QUICK START
git clone https://github.com/chutingyean/Agent-Cockpit.git
cd Agent-Cockpit
npm install
npm run demo
node bin/cockpit.mjs serve --repo demo/research
node bin/cockpit.mjs demo --breakCommands are taken from the repository README — check the current version on GitHub before running.
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram