Autoresearch

Autoresearch is an open-source tool designed to automate goal-directed iteration for Claude Code across a wide range of domains. Inspired by Karpathy's autoresearch principles, it enables autonomous modification, verification, and management of changes in code, machine learning projects, content, and performance tasks. The tool operates by having Claude make one atomic change per iteration, verify the change using a measurable metric, and then either keep or revert the change automatically, ensuring a clear and traceable history of what modifications led to improvements.

Autoresearch provides several features to streamline autonomous iteration. It uses mechanical verification, where each iteration is evaluated based on real metrics such as test coverage, build time, or performance benchmarks, rather than subjective assessments. Automatic git rollback is implemented to instantly revert failed changes, eliminating the need for manual cleanup or debugging of compound failures. The tool logs every iteration in a TSV file, recording metrics, deltas, statuses, and descriptions, which allows for pattern recognition across experiments. Git is used to track commits, and the agent references its own history to avoid repeating unsuccessful modifications.

The platform includes a planning wizard that guides users through defining their goals, suggesting appropriate metrics, validating verification commands with a dry run, and outputting a ready-to-launch configuration. Security auditing is also integrated, featuring automated STRIDE threat modeling, OWASP Top 10 coverage, and red-team persona testing. The security audit generates structured reports, maps attack surfaces, and provides prioritized mitigations. Additional flags allow for delta audits, auto-fixing critical or high findings, and enforcing CI/CD severity gates. Each audit produces a timestamped report folder with comprehensive documentation, including threat models, findings, and recommendations.

Autoresearch is domain-agnostic and can be applied to backend code, frontend UI, machine learning training, performance optimization, refactoring, and content quality improvements, as long as a measurable outcome is defined. Installation involves cloning the repository and copying the skill to the appropriate directory for Claude.