ClaudeMap

Agent Cockpit

A local whiteboard plus an MCP server for AI research agents: plans, tasks, runs and claims land on one board, and when any link in an evidence chain changes, the claims that depend on it turn red — deterministic rule checks, zero LLM calls (AGPL-3.0)

Plugins & Agentsresearchevidence chainmcpwhiteboardcode review

Agent Cockpit is a local evidence-chain workbench for research done with AI agents (npm package @tingyen/agent-cockpit, needs Node 22.13+ for the built-in node:sqlite; 465 stars as of writing, 2026-11-09). Any MCP-speaking agent (Claude Code, Codex, Cursor, Gemini CLI) draws its work onto the same board: the research plan and stages, task cards, every experiment run (recording the script's hash, every file it depends on, and per-function reachability hashes) and every claim (which must cite a file and field a recorded run actually wrote). A deterministic engine keeps checking the whole chain: change any link's hash and the run goes stale, with claims built on it flagged for recheck; revert the file, hashes match again and the claims recover — staleness is decided by content hashes, not timestamps, and for Python dependencies precision goes down to the function (a change to a function the run never reaches does not stale it). Unreported changes are caught whether or not the agent admits them: they block git commit, and Claude Code cannot end its turn until it reports. Every write (MCP, CLI, board) passes one validation gate, and non-compliant writes are rejected with a reason and logged. Questions for a human go through cockpit_decide, and when a change breaks claims the board raises the decision itself with a one-page evidence brief. Everything is appended to .cockpit/events.jsonl as the single source of truth (the author suggests committing it), the board is a pure projection, and you can replay to any moment. A Claude Code plugin ships (with a skill that teaches agents how to work under Cockpit), and the MCP server is listed in the MCP Registry as io.github.chutingyean/agent-cockpit. The board binds to 127.0.0.1, and every check is a plain rule with zero tokens.

What it does

  • Evidence chains down to a JSON field: claims must cite values a recorded run wrote, and runs record hashes of the script, its file dependencies and every reachable function
  • Unreported changes get caught whether or not the agent reports them — they block git commit and Claude Code cannot end its turn until it reports
  • One write gate for MCP, CLI and board: non-compliant writes are rejected with a reason the agent can act on, and logged
  • Human decisions in the loop: agents ask via cockpit_decide and wait; broken chains raise a decision with a one-page evidence brief automatically
  • Append-only event store (.cockpit/events.jsonl) as the single source of truth — the board is a projection, so you can scrub or replay to any moment
  • Ships a built-in graphify-style code graph and rule-based reviewer; install the real graphify or alibaba/open-code-review and they are used automatically

When to use it

For long-running research-flavored agent work — experiments, graders and data pipelines that interlock, where one agent quietly editing shared code silently invalidates other "verified" results, which is exactly the failure mode Cockpit watches for with machine rules; also for teams who want agents to stop and ask a human at decision points.

Getting started

# on your own research repo, no clone needed (Node 22.13+, git; python3 for the demo):
npx -y @tingyen/agent-cockpit init      # MCP configs, AGENTS.md, Stop hook, pre-commit
npx -y @tingyen/agent-cockpit serve     # the board at http://127.0.0.1:4180
npx -y @tingyen/agent-cockpit run -- python3 train.py   # run experiments through the cockpit

# permanent command:
npm install -g @tingyen/agent-cockpit

# Claude Code plugin (adds a skill that teaches agents to work under Cockpit):
claude plugin marketplace add chutingyean/Agent-Cockpit
claude plugin install agent-cockpit@agent-cockpit

# try the demo first:
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

Things to know

License first: AGPL-3.0-or-later — if you modify it and serve it over a network, you must offer the source of your version. Local by design: the board binds to 127.0.0.1 and checks are plain rules, not an LLM judge. Function-level staleness precision applies to Python dependencies; other languages fall back to file-level hashes. The built-in code graph and reviewer are deliberately lightweight stand-ins — the real graphify and open-code-review are used automatically when installed. Committing .cockpit/events.jsonl is part of the design (it doubles as the agents' work log). Node 22.13+ is required for the built-in node:sqlite. 465 stars as of writing (2026-11-09).

Repository

chutingyean/Agent-Cockpit

Charted

Related on the map

  • clawd-conduit

    Plugin / Agent

    Wire all 15 Claude Code hook events into the Clawd on Desk desktop pet — and turn permission requests into a blocking desktop card.

    • hooks
    • desktop
    • notifications
    • claude-code
  • CometixCode

    Plugin / Agent

    An unofficial Rust 1:1 reimplementation of the Claude Code terminal UI, ported file-for-file onto an iocraft TUI framework.

    • terminal-ui
    • rust
    • reimplementation
    • unofficial
  • Abide

    Plugin / Agent

    Hook-level enforcement for the natural-language rules in AGENTS.md and CLAUDE.md — every edit is judged and violations are repaired in the same turn.

    • hooks
    • code-review
    • rules
    • agents-md