CatoBot Autoexperiment MCP Server

Example Use Cases

• 1

  Autonomously modify code or configurations and evaluate results in a repeatable experiment loop.

• 2

  A standardized interface to run experiments, track results, and manage experiment state across diverse domains.

• 3

  Integration with git to version control experiment changes and roll back regressions automatically.

Description

This MCP server generalizes Karpathy's autoresearch pattern into a reusable Model Context Protocol server that any AI agent can drive for autonomous experimentation across any domain. It manages experiment lifecycle including file edits, running experiments, measuring results, and deciding to keep or discard changes, all configured via a JSON file. The server supports shell, external, and hybrid execution modes, integrates with git for version control, and provides a consistent interface for agents to optimize a single metric efficiently.