Hackable terminal coding agent. Minimal core, extensible skills, your way. A compact agent loop with local tools, context compression, and multi-provider support.
Intentionally small codebase. The core runtime is straightforward — agent loop, tools, context compression, CLI. Nothing extra.
Extend with skills, swap models, customize tools. The architecture is built to be modified, not just used.
Drop-in skill system. Write a SKILL.md, add files, and your agent learns new capabilities instantly.
Works with both OpenAI-compatible and Anthropic APIs. Switch providers without changing your workflow.
Minimal agent loop that supports parallel execution for multiple tool calls, keeping responses fast.
Scan ~/.kittycode/skills for skill folders. Each SKILL.md is automatically discovered and injected into the agent's context.
Shell execution, file operations, web search, web fetch/crawling, TODO tracking, and sub-agents — all included.
Save and resume sessions across terminal restarts. Context compression keeps long sessions manageable.
Press Esc to interrupt at the next safe checkpoint between LLM streaming, tool dispatch, and loop rounds.
List all loaded skills discovered at startup.
Switch between configured model providers on the fly.
Manually trigger context compression for long sessions.
Save current state and browse previous sessions.