Autoresearch
Claude Autoresearch Skill — Autonomous goal-directed iteration for Claude Code. Inspired by Karpathy's autoresearch. Modify → Verify → Keep/Discard → Repeat forever.
Last verified:
What is Autoresearch?
Autoresearch is an open-source Claude Code skill that enables autonomous goal-directed iteration for any task with a measurable outcome. Inspired by Andrej Karpathy's autoresearch concept, it lets Claude loop autonomously—making one atomic change per iteration, committing before verification, and automatically reverting if the metric doesn't improve. The tool follows a modify → verify → keep/discard → repeat cycle without human intervention between iterations.
Key features include automatic git rollback on failures, mechanical verification using real metrics (test coverage, build time, Lighthouse score, val_bpb), a plan wizard for interactive goal-to-config setup, and a new v1.0.3 security audit mode with STRIDE threat modeling plus OWASP Top 10 checks. Results are logged in TSV files while kept changes become git commits, giving you a clean history of what worked. The latest version adds four red-team personas, auto-fix for Critical/High findings, and CI/CD severity gates.
Autoresearch is designed for developers, researchers, and engineers who want to ship 10x more output by running autonomous iteration loops while sleeping. It works across backend code, ML training, frontend UI, performance optimization, refactoring, content creation, and security audits. The skill is domain-agnostic—any task with a measurable outcome can benefit from autonomous looping.
Autoresearch pricing
Pricing model: Freemium
Free. Open source under MIT license. No paid tiers or subscriptions. Two commands to install, one command to run. Available across 6 platforms including Claude Code, Codex, Copilot, Cursor, Gemini CLI, and Windsurf.
Autoresearch pros
- One atomic change per iteration—exact cause of failures is clear
- Automatic git rollback on failures—no manual cleanup needed
- Works on any domain—code, ML, content, performance, refactoring
- Mechanical verification with real metrics, not subjective judgment
- Commit before verify ensures clean git history
- Auto-revert on failure saves time debugging compound failures
- Git as memory—agent learns from its own history what works
- TSV results logging tracks every iteration with metric and delta
- Plan wizard converts plain-language goals to ready-to-launch config
- Security audit with STRIDE + OWASP Top 10 + 4 red-team personas
- Auto-fix confirmed Critical/High security findings automatically
- Bounded iterations with /loop N for controlled runs
- Early completion when goal achieved before N iterations
- CI/CD severity gate with --fail-on flag for pipeline blocking
- Free and open source under MIT license
Autoresearch cons
- Requires Claude Code installation to use
- Needs mechanical metric that can be verified programmatically
- Only works on tasks with measurable outcomes
- Must have git repository initialized in project directory
- Security audit v1.0.3 is new and may have undiscovered bugs
- Unlimited loops may run indefinitely without /loop N bound
- Requires custom verify command setup for each use case
- TSV logging may require custom parsing for analysis
Frequently asked questions about Autoresearch
What is Autoresearch?
Autoresearch is an open-source Claude Code skill that enables autonomous goal-directed iteration for any task with a measurable outcome. Inspired by Karpathy's autoresearch, it lets Claude loop autonomously—modify one atomic change, commit, verify the metric, keep or revert, then repeat forever without human input.
How do I install Autoresearch?
Run two commands: git clone https://github.com/uditgoenka/autoresearch.git /tmp/autoresearch, then cp -r /tmp/autoresearch/skills/autoresearch ~/.claude/skills/autoresearch. Then type /autoresearch inside any project directory to start the autonomous loop.
What metrics can I use?
Any mechanical metric that is measurable: test coverage percentage, build time in seconds, Lighthouse score, val_bpb for ML training, benchmark time in milliseconds, LOC reduced for refactoring, word count + readability for content. The verify command must be programmatically executable.
What happens if an iteration fails?
Failed changes revert instantly via git reset. No manual cleanup is needed and there is no debugging of compound failures since only one atomic change was made per iteration. The rejection is logged in the TSV file.
Can I limit the number of iterations?
Yes. Use /loop N /autoresearch to run exactly N iterations then stop with a summary. Early completion happens when the goal is achieved before N iterations. Final summary shows baseline → current best, keeps/discards/crashes.
What is the security audit feature?
v1.0.3 adds autonomous STRIDE threat model + OWASP Top 10 (70+ checks) + 4 red-team personas. It generates a timestamped folder with 7 markdown files: overview, threat model, attack surface, findings, OWASP coverage, dependency audit, and recommendations. Flags include --diff for delta mode, --fix for auto-remediate, --fail-on for CI gate.
What is the plan wizard?
/autoresearch:plan is an interactive wizard introduced in v1.0.2 that converts your plain-language goal into Scope, Metric, and Verify config. It scans your codebase, suggests metrics by domain, dry-runs the verify command to validate it works, and outputs a ready-to-paste config.
What domains does Autoresearch work on?
It is domain-agnostic: backend code (test coverage), frontend UI (Lighthouse score), ML training (val_bpb/loss), performance (benchmark time), refactoring (LOC reduced + tests pass), blog/content (word count + readability), and security audit (OWASP + STRIDE coverage).
How does git as memory work?
Every kept change becomes a git commit. The agent reads its own git history to learn what works and avoid past mistakes. This creates a clean history documenting what succeeded and what failed across iterations.
Is Autoresearch free?
Yes, it is completely free and open source under MIT license. There are no paid tiers, subscriptions, or hidden costs. Two commands to install, one command to run, and it is available across 6 platforms.