Autoresearch

Autoresearch

Domain-agnostic autonomous experiment runner. Optimizes a single metric by iterating: hypothesis -> code change -> verify -> keep/discard.

What You Get

• JSONL experiment log with metric history, deltas from baseline, and commit SHAs
• Markdown results dashboard comparing all iterations
• Git branch per experiment with keep/discard via commit/revert

CLI Usage

# Initialize an experiment
autoresearch init noir-gates \
  --objective "minimize constraint count" \
  --metric constraints \
  --verify "nargo compile 2>&1 | grep constraints" \
  --direction lower \
  --mutable src/main.nr

# Run a single iteration
autoresearch run "inline witness computation"

# Autonomous loop (Claude generates hypotheses)
autoresearch loop --iterations 10 --model claude-sonnet-4-6

# View results
autoresearch dashboard
autoresearch status

MCP Server

autoresearch serve

Configure MCP

Add to ~/.mcp.json:

{
  "mcpServers": {
    "autoresearch": {
      "command": "autoresearch",
      "args": ["serve"]
    }
  }
}

MCP Tools

| Tool | Description | |------|-------------| | autoresearch_status | Current experiment status (name, metric, runs, best) | | autoresearch_dashboard | Full results table as markdown | | autoresearch_run | Execute a single experiment iteration |

Install

cd agents/autoresearch
pip install -e .

Requires ANTHROPIC_API_KEY for the autonomous loop (hypothesis generation).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: DROOdotFOO
• Source: DROOdotFOO/agent-skills
• License: MIT
• Homepage: https://droo.foo/

Install and usage instructions live in the source repository linked above.