yologdev/yoyo-evolve

What it solves

yoyo is a coding agent that works in your terminal, helping you edit code, run commands, manage git, and more. But its most unusual feature is that it also rewrites its own source code: it reads its own codebase, decides what to improve, implements the change, runs tests, and commits—all without a human. This lets the project grow itself over time.

How it works

Every few hours yoyo wakes up and follows a routine: it reads its own source files, checks GitHub issues for community suggestions, plans an improvement, makes changes, runs the test suite, and only commits if tests pass (otherwise it reverts). It also has "social sessions" where it reads GitHub Discussions and replies to conversations, and a daily job that compresses its accumulated learnings into active memory files that get loaded into every prompt. Under the hood, yoyo is a normal LLM-powered agent: it uses APIs from Anthropic, OpenAI, Google, and others, can spawn sub-agents for parallel tasks, and offers a rich set of slash commands for editing, git, code review, project health, and security scanning.

Who it's for

Developers who want an AI coding assistant that can be customized and that also can improve itself. The project is open source, and the community can influence yoyo's evolution by filing GitHub issues labeled agent-input; the agent prioritizes issues with more thumbs-up reactions.

Highlights

• Fully autonomous self-evolution: yoyo reads, modifies, tests, and commits changes to its own codebase every few hours, growing from 200 lines to over 158,000 lines in 180 days.
• 100+ slash commands in a streaming REPL: code editing, git operations, PR workflows, project health checks, linting, refactoring, session management, and more.
• Multi-provider support: works with Anthropic, OpenAI, Google, Ollama, OpenRouter, and many other LLM providers, switchable mid-session.
• Extensible: custom slash commands, MCP servers, OpenAPI tools, and a skills system; also supports sub-agents with parallel and background execution.
• Strong safety features: permission system with allow/deny lists, path traversal prevention, and a trust boundary that refuses project-level configs from silently granting privileges.