Agent Cockpit
A local whiteboard and MCP server that traces every AI research agent claim back to the runs and code that produced it.
At a Glance
Self-hosted local whiteboard and MCP server, free software under the GNU Affero General Public License v3 or later. Runs locally, binding to 127.0.0.1, with no LLM calls.
Engagement
Available On
Listed Nov 2026
About Agent Cockpit
Agent Cockpit is an open-source local whiteboard and MCP server from chutingyean for AI agents that do research. Agents such as Claude Code, Codex, Cursor and Gemini CLI draw their plan, tasks, experiment runs, claims and dependent code onto one board. A deterministic engine keeps checking the evidence chain and marks dependent cards red when a link changes.
What It Is
Agent Cockpit is an evidence-chain tracker for agent-driven research repositories. Any MCP-capable agent writes to the same board through a single validated write gate. A claim must cite a value that a recorded run actually wrote, so each result stays tied to the script and code behind it.
How the evidence chain works
Each run records hashes of its script, of every file the script depends on, and of each function it reaches. Staleness is decided by comparing content hashes rather than timestamps, so reverting a change restores validity. For Python dependencies, changes to functions a run never reaches do not make it stale. The state is stored as an append-only event log in .cockpit/events.jsonl, which allows the board to be replayed to any moment.
Checks, review and human decisions
Real file changes are detected whether or not an agent reports them, and unreported changes block git commit. A built-in call graph supplies dependency closures and impact analysis, and a rule-based reviewer checks changed lines. When a change breaks claims, the cockpit raises a decision and writes a one-page evidence brief for a human to answer. The checks are plain rules with no LLM calls, and the board binds to 127.0.0.1.
Setup path
It requires Node.js 22.13+ and git. Users run cockpit init in a git repo to write MCP configs, an AGENTS.md, a Stop hook and a pre-commit hook, then cockpit serve to open the board. A Claude Code plugin and a skill file are also provided, and the MCP server is listed in the MCP Registry.
Community Discussions
Be the first to start a conversation about Agent Cockpit
Share your experience with Agent Cockpit, ask questions, or help others learn from your insights.
Pricing
Open Source (AGPL-3.0-or-later)
Self-hosted local whiteboard and MCP server, free software under the GNU Affero General Public License v3 or later. Runs locally, binding to 127.0.0.1, with no LLM calls.
- Licensed under AGPL-3.0-or-later (OSI-approved)
- Local whiteboard plus MCP server for AI research agents
- Evidence chains with content-hash staleness checks
- Single write gate for MCP, CLI and board
- Built-in code graph and rule-based code review
Capabilities
Key Features
- Evidence chains linking claims to recorded runs and code hashes
- Function-level staleness detection for Python dependencies
- Single validated write gate for MCP, CLI and board writes
- Detection of unreported file changes with git pre-commit and Claude Code Stop hook
- Built-in code graph and rule-based code review
- Human decision requests with one-page evidence briefs
- Append-only event log with board replay
- Local deterministic checks with no LLM calls
- MCP tools for plans, stages, tasks, runs, claims, hypotheses and decisions
- Claude Code plugin and SKILL.md
