SkillFoxx

Agent Cockpit MCP

A local whiteboard and MCP server for research agents: every claim is tied to the run and the code that produced it

  • 480 ★ on the repository
  • Last commit
  • AGPL-3.0
MCP server

Extended access

Why this level

  • The `cockpit_run` tool runs commands on your machine and records the result
  • `cockpit init` adds a Stop hook and a pre-commit hook that block commits and the agent's turn from ending
  • The board listens on `127.0.0.1`, but the event log in the repository stores commands and claims
All reasons and checks

chutingyean/agent-cockpitAuthor:chutingyean

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add mcp/agent-cockpit

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • The `cockpit_run` tool runs commands on your machine and records the result
  • `cockpit init` adds a Stop hook and a pre-commit hook that block commits and the agent's turn from ending
  • The board listens on `127.0.0.1`, but the event log in the repository stores commands and claims

This is third-party code. Review the repository files before installing.

What it does

Agent Cockpit draws the work of research agents on one local whiteboard: the plan, the tasks, every experiment run, the claims and the code they depend on. The board opens at `http://127.0.0.1:4180`. An MCP-capable agent (Claude Code, Codex, Cursor, Gemini CLI) writes to it through tools such as `cockpit_plan`, `cockpit_task`, `cockpit_run` and `cockpit_claim`. The core is the evidence chain. A claim must cite a value in a file written by a recorded run. A run stores hashes of its script, the files it depends on and the Python functions it reaches. If an agent edits the grader or the data pipeline, the run goes stale and every claim built on it turns red. Reverting the change restores the hashes and the claims count as valid again. All checking is deterministic and makes no language model calls. A single write gate rejects writes that break the rules and gives a reason. Unreported file changes are caught: they block `git commit`, and Claude Code cannot end its turn until it reports them. When a change breaks claims, the board raises a question for a human and writes a one-page brief. History is stored as events in `.cockpit/events.jsonl` and can be replayed.

Who it is for. Machine learning researchers and analysts who hand experiments to agents and want to see which results still rest on unchanged code.

Good fit when

  • An agent ran training for a week and you need to know which claims survived its edit of the grader
  • Two agents, Claude Code and Codex, work on one study and have to write to the same board
  • An agent keeps tweaking hyperparameters and you need a point where a human decides through `cockpit_decide`
  • You need a reproducible log: which script, on which files and with which hashes produced a figure in the report

Not a fit when

  • The project is not in a git repository: `cockpit init` works only inside one
  • You need a shared web service for a team: the board listens only on `127.0.0.1` and is local
  • The code is not Python and you need function-level precision: that precision applies to Python dependencies

Example request

Use cockpit to check which of my claims went stale after the edit to eval/judge.py and show which run needs to be repeated

MCP

Transport
stdio
Authentication
not required

Limitations

  • Node.js 22.13 or newer (for the built-in `node:sqlite`) and git are required, plus `python3` for the demo project.
  • The package is not published to npm: installation is `git clone` and `npm install`, and the `cockpit` command appears after `npm link`.
  • It is version 0.1.0 under AGPL-3.0 or later: if you modify it and let others use it over a network, you must release the source.
  • A fuller code graph and code review turn on if you install graphify and open-code-review; without them lightweight built-in versions run.

How to disable. Remove the `cockpit` entry from the client's MCP config, remove the Stop hook and the pre-commit hook that `cockpit init` wrote, and delete the `.cockpit/` folder if you do not need the event log.

FAQ

How do I install Agent Cockpit?

Run `git clone https://github.com/chutingyean/Agent-Cockpit.git`, then `npm install` in the repository folder. In your own research repo run `node /path/to/Agent-Cockpit/bin/cockpit.mjs init` and open the board with `serve`.

What does `cockpit init` write?

MCP configs for Claude Code, Cursor, VS Code and Gemini CLI, an `AGENTS.md` file, a Stop hook and a pre-commit hook. For Codex it prints the connect command. It writes only inside your repository.

How do I tell that a claim no longer holds?

The claim's card on the board turns red. The reason sits beside it: which run went stale and which file or function changed. Reverting the edit restores the hashes and the claim turns green again.

Can I try it without my own project?

Yes. `npm run demo` builds a demo repository with six runs and six claims, and `node bin/cockpit.mjs demo --break` quietly loosens the grader to show claims turning red.

Does checking use language model tokens?

No. Every check is a plain rule, so no tokens are spent. The agent that writes to the board uses its own usual subscription or key.

Discussion

No questions or reviews yet.

MCP servers

A shared board for Claude Code, Codex and other agents: a question an agent stops on rises to the top within seconds.

PluginExtended access58

MCP server for a shared agent room: decisions, tasks verified by another agent and a final report, joined by code.

MCP serverExtended access107

A CLI that reads local Claude Code, Codex and Hermes sessions, finds repeated failures and stages fixes for manual review

CLIExtended access73

A pixel-art office in the browser: each Claude Code or Codex session sits at a desk, types and raises a hand when it needs you.

PluginExtended access60

A bridge between local CLI agents: Codex, Cursor or Claude Code delegate work to Antigravity, Grok, Kimi and seven more.

MCP serverBroad access131

Local meeting rooms for Codex and Claude Code agents to talk to each other while a person reads and posts from a browser.

SkillExtended access113

README in short

The README opens with the problems of long agent research: claims going stale silently, unreported edits and unanswered questions to a human. It then covers features, a quick start with a demo, connecting your own repository, how the hashes work, tables of MCP tools and CLI commands, and the project layout.

README badge

Report a problem