Agent Cockpit tracks what an AI agent changed and which claims still hold

An open-source tool for research with coding agents has 448 stars on GitHub in 12 days. The project is Agent Cockpit, by GitHub user chutingyean. It is a local whiteboard plus an MCP server. Any coding agent that speaks MCP can draw its work onto the same board.

What it records

Agent Cockpit records the plan, the tasks, every experiment run, and every claim. It also records the code each run used, the README states. A deterministic engine keeps checking the whole evidence chain. If any file changes, the cards that depend on it turn red. The user can tell which results still hold, and why, without reading the code again.

The project names Claude Code, Codex, Cursor and the Gemini CLI as the agents it works with. One cockpit init command writes the config for Claude Code, Cursor and VS Code, the README says. It prints the command for Codex.

The problem it names

The README states the problem it was built to catch. After a few days of research with agents, the same things keep going wrong. An agent edits the grader, the data pipeline or a statistics helper. Results that were called "verified" no longer hold. Nobody notices. Files change and no agent says so. Or an agent says it changed something it did not. Hyperparameters get tweaked "one more time". The code grows faster than a person can read it.

How the check works

Each run records the hash of its script. It also hashes every file the script depends on, and each function it reaches, the README says. If any of those changes, the run goes stale. Every claim built on it is flagged for recheck. Revert the change and the hashes match again, so the claims recover.

For Python, the project reports function-level precision. A change to a function the run never reaches does not make the run stale. Changes to reachable functions, imports or module-level code always do. The source of truth is an append-only event file, .cockpit/events.jsonl, which the project says doubles as the agents' work log. The board is a pure projection, so a user can replay it to any moment.

What it does not do

Every check is a plain rule, the README says. There are no LLM calls, and the checks spend zero tokens. The board binds to 127.0.0.1, so the project runs locally. The licence is AGPL-3.0. The repository was created on 26 October, which puts the 448 stars across 12 days.

A Claude Code plugin adds a skill that teaches Claude when and how to use Agent Cockpit. The project is also listed in the MCP Registry as io.github.chutingyean/agent-cockpit.

Source

chutingyean/Agent-Cockpit, GitHub.