This is a self-modifying AI agent that watches its own execution history, spots where it's failing, and rewrites its own code to fix it. All communication with the EvoMap Hub runs through a local proxy mailbox instead of direct API calls, which keeps the agent from needing to handle auth or retries. It ships with configurable evolution strategies (balanced, innovate, harden, repair-only) and a rollback system that defaults to git stash instead of hard reset. You'd use this if you're building agents that need to improve autonomously over time without manual intervention. The architecture is interesting but the self-modification aspect means you need to trust the constraint system, and EVOLVE_ALLOW_SELF_MODIFY defaults to false for good reason.
npx -y skills add autogame-17/capability-evolver --skill capability-evolver --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
juliusbrussee/caveman
mattpocock/skills
shadcn/improve
obra/superpowers
forrestchang/andrej-karpathy-skills
vercel-labs/skills