I've built kit originally as an experiment. Goal was to find out how context efficient I can make a headless agent. It worked so well that I ended up building a TUI and replacing Claude Code, Codex and OpenCode with kit as my daily driver.
Standout features:
- dynamic reload of skills, plugins, MCP servers, no more /reload-plugins, /reload-skills or exit --resume - only a single meta tool: compose - it uses my custom language, Runlet, which is optimised for tool call composition for LLMs - first-class ACP support - use kit inside any ACP client or launch any ACP harness inside kit as a subagent (supports Claude, Codex, Cursor and more) - A2A out-of-the-box: launch kit on a machine, have other agents talk to it
Works with your ChatGPT subscription usage limits, OpenRouter or Speakeasy sign in.
Let me know what you think & AMA