Autoresearch is a shell-based tool to run autonomous goal-directed iteration across code or other domains, using a fixed loop of modify–verify–keep/discard with an orchestrator. It supports Claude Code, OpenCode, and Codex, with 14 commands and safety hooks.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Autoresearch is a skill for Claude Code, OpenCode, and Codex that enables autonomous, iterative improvements driven by a plain-language goal, a measurable metric, and a loop that modifies code, verifies changes, and keeps or discards improvements. It follows a structured loop inspired by Karpathy's autoresearch, with a focus on constrained scope and automated rollbacks.
How it works
The tool operates a LOOP for N iterations or until a goal is met:
- Review current state, git history, and results log
- Pick the next change based on prior outcomes
- Make one focused change
- Git commit before verification
- Run mechanical verification (tests, benchmarks, scores)
- If improved, keep; if worse, revert; if crashed, fix or skip
- Log the result
- Repeat until completion
Setup phase before looping:
- Read context
- Define goal (mechanical metric)
- Define scope (modifiable vs read-only files)
- Establish baseline by running verification on the current state
- Confirm and start the loop
Key rules include bounded iterations by default, read-before-write, automatic rollback, and git history memory via the experiment prefix.
Getting started
Quick-start options include:
- Claude Code:
npx skills add uditgoenka/autoresearch
- Plugin install:
/plugin marketplace add uditgoenka/autoresearch
/plugin install autoresearch@autoresearch
- Manual copy:
git clone https://github.com/uditgoenka/autoresearch.git
# Copy skill + subcommands to your project
cp -r autoresearch/.claude/skills/autoresearch .claude/skills/autoresearch
cp -r autoresearch/.claude/commands/autoresearch .claude/commands/autoresearch
cp autoresearch/.claude/commands/autoresearch.md .claude/commands/autoresearch.md
- Updating (no reinstall needed):
/plugin update autoresearch
- OpenCode, Codex quick starts are provided similarly in the README.
Run the main loop with an example configuration:
/autoresearch
Goal: Increase test coverage from 72% to 90%
Scope: src/**/*.test.ts, src/**/*.ts
Metric: coverage % (higher is better)
Verify: npm test -- --coverage | grep "All files"
Iterations: 25
Releases
Latest versions include v2.2.1 (Orchestrator Seam Hardening) and v2.2.0 (Autonomous Goal-directed Orchestrator), with prior updates focusing on regression gates and knowledge base improvements. The latest release date listed is 2026-06-23 for v2.2.1. Features include automatic orchestration, plain-language goal handling, and iterative subcommand selection.
Traction
Stars: 5753, Forks: 436, Open issues: 6
Behind the repo
No linked startup/company information is provided beyond the project maintainers and the repository itself.
Caveats
License: MIT. Created 2026-03-13; last push 2026-06-23. Topics include ai, autonomous-agent, autoresearch, claude-code, iteration, and related domains.