#9797th of 149 in Skills, plugins and rules
autoresearch
Skill that runs coding agents in a measured keep-or-revert loop
- Stars
- 6.6k
- License
- MIT
- Last commit
- Aug 2026
- Last release
- Aug 2026
- Language
- Shell
2,288 on 25 Mar 20266,574 on 14 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Autoresearch is a skill pack for Claude Code, OpenCode and OpenAI Codex, based on Karpathy's autoresearch. Given a goal, a metric and a verify command, the agent makes one change per iteration, commits it, runs the check, and keeps or reverts the commit. It adds 14 subcommands (plan, debug, fix, security, ship, scenario, regression and others) and, for Claude Code only, hook guardrails.
Who it is for: Developers using Claude Code, OpenCode or Codex who want metric-driven loops
Strengths
- Works with Claude Code, OpenCode and Codex; installs via npx skills, plugin or install.sh
- Each iteration is a git commit with automatic revert on failure and a TSV results log
- Every command has a default iteration bound; unlimited runs are opt-in
- Claude Code hooks block force-push, rm -rf, .env reads and node_modules context bloat
Weaknesses
- Hook guardrails exist only for Claude Code, not OpenCode or Codex
- Classic loop needs a verify command that outputs a number
- README calls the hooks defense-in-depth, not a security sandbox
- Autonomous commit and revert loop on your repo; token spend scales with iterations
What it needs
- no GPU
- Needs Claude Code, OpenCode or OpenAI Codex, git
- Models: Claude Code, OpenCode, OpenAI Codex
Also in Skills, plugins and rules
See all 149| Rank | Project | Score |
|---|---|---|
| 1 | cavemanCuts token use in coding agents with terse replies and compressed inputs | 95 out of 100 |
| 2 | skillsAgent skills for planning, TDD, debugging and code review | 94 out of 100 |
| 3 | eccAgents, skills, hooks and rules bundle for Claude Code, Codex and other harnesses | 94 out of 100 |
| #4 | impeccableDesign skill with 24 commands and a detector for AI-generated frontends | 94 out of 100 |
| #5 | ponytailSkill prompt that makes coding agents write less code and add tests | 92 out of 100 |