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.
Hello HN,
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