Self-Improving AI Agent
A terminal coding agent that runs a real think→tool→result loop (read / edit / write / shell / git / search / web, MCP tools in-loop, confirm + undo) and rewrites its own system prompt from feedback, with metrics and automatic rollback — free by default via a local model router.
Self-Improving AI Agent
A terminal coding agent that executes tools and rewrites its own system prompt from your feedback, with per-version metrics and automatic rollback.
What's inside
- Real
think → tool → resultloop: read / edit / write / shell / git / search / web / worktree / notebook - Confirmation for writes and commands, working undo, shell-injection and SSRF guards
- MCP tools callable mid-conversation; sub-agents (reviewer, test writer, debugger, researcher, refactorer)
- Feedback pipeline ships an improved prompt; auto-rollback if it underperforms
- Free by default via the FCM local model router, or 6 keyed providers / a Claude subscription
Use it for
A self-hosted coding agent that adapts its behaviour to how you like to work.