Repolode
Submit
uditgoenka

uditgoenka/autoresearch

A Claude Code, OpenCode and Codex skill that runs goal-driven modify-verify-keep/discard loops, inspired by Karpathy's autoresearch.

View on GitHubWebsite
Stars
6,582
Forks
482
Open issues
8
Last commit
2 months ago

What is it?

Autoresearch is a skill pack for AI coding agents such as Claude Code, OpenCode and OpenAI Codex. You give it a goal, a measurable metric and a verify command, and the agent repeatedly makes one focused change, commits it, runs the verification, keeps the change if the metric improves and reverts it otherwise. Each iteration is logged in TSV, and git history serves as the agent's memory between rounds.

Beyond the core loop, it bundles 14 commands for specific jobs: planning a metric from a fuzzy goal, hunting and fixing bugs, running STRIDE and OWASP security audits, generating edge-case scenarios, multi-persona prediction, documentation generation, adversarial reasoning over subjective decisions, requirement probing, product improvement research, results analysis and a regression stability gate. An optional Guard command protects existing tests while a metric is being optimized.

The project is built on Karpathy's autoresearch idea and generalizes it beyond machine learning to code, content and other domains that have a number to measure. Hook guardrails that block risky commands or sensitive files are available for Claude Code only.

Key features

  • 01Runs a modify, verify, keep-or-revert loop with one change per iteration.
  • 02Provides 14 commands for planning, debugging, fixing, security audits and more.
  • 03Logs every iteration in TSV format and analyzes trends across runs.
  • 04Supports Claude Code, OpenCode and OpenAI Codex with installer scripts.
  • 05Includes a Guard option to protect existing tests during optimization.
  • 06Defaults to bounded iteration counts, with unlimited mode opt-in.

Use cases

  • Raising test coverage or reducing bundle size with an agent running unattended overnight.
  • Auditing a codebase for STRIDE and OWASP risks with a structured report.
  • Fixing a backlog of failing tests, type errors or lint errors until the count reaches zero.
  • Stress-testing requirements or design decisions before writing code.

Quick start

  1. 1

    Install the skill for Claude Code via npx

    npx skills add uditgoenka/autoresearch
  2. 2

    Start a new Claude Code session so the commands load

  3. 3

    Run the autoresearch command with a goal

    /autoresearch

Things to know

  • Hook guardrails work only in Claude Code; OpenCode and Codex lack hook parity.
  • Loops need a mechanical metric or verification command to measure progress.
  • Claude Code requires a new session after installation before reference files resolve.

Topics

#claude-code#ai-agent#autonomous-agent#iteration#llm-tools#code-quality#security-audit#ai#autoresearch#claude

FAQ

What is autoresearch?
A Claude Code, OpenCode and Codex skill that runs goal-driven modify-verify-keep/discard loops, inspired by Karpathy's autoresearch. Autoresearch is a skill pack for AI coding agents such as Claude Code, OpenCode and OpenAI Codex. You give it a goal, a measurable metric and a verify command, and the agent repeatedly makes one focused change, commits it, runs the verification, keeps the change if the metric improves and reverts it otherwise. Each iteration is logged in TSV, and git history serves as the agent's memory between rounds.
Is autoresearch free to use? What license does it have?
autoresearch is open source under the MIT license. Check the license terms before using it in commercial products.
Is autoresearch still maintained?
Updated within 6 months. The last commit to autoresearch was on Aug 12, 2026.
What can I use autoresearch for?
Raising test coverage or reducing bundle size with an agent running unattended overnight. Auditing a codebase for STRIDE and OWASP risks with a structured report. Fixing a backlog of failing tests, type errors or lint errors until the count reaches zero. Stress-testing requirements or design decisions before writing code.
How do I install autoresearch?
Install the skill for Claude Code via npx: npx skills add uditgoenka/autoresearch Start a new Claude Code session so the commands load Run the autoresearch command with a goal: /autoresearch
Where can I get autoresearch?
The source code and installation instructions are on GitHub: https://github.com/uditgoenka/autoresearch

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects