agenthatch
A tool that compiles a SKILL.md file into a standalone, runnable Python AI agent instead of loading it as a prompt, addressing context isolation, validation and scaling problems with treating skills as loose instructions.
agenthatch treats a SKILL.md file as source code rather than a prompt that an agent interprets at runtime. It parses the skill, runs it through a multi-stage large-language-model pipeline, and generates code with a templating engine, producing a self-contained Python package with its own pyproject.toml, a command-line entry point, typed tool definitions, Model Context Protocol integration and a runtime configuration.
The stated goal is to fix problems the author attributes to markdown-based skills used across tools such as Claude Code, Codex CLI and OpenClaw: skills leaking context into each other, agents skimming rather than following long instructions closely, every loaded skill consuming context-window tokens even when unused, no validation of tool names or parameters before runtime, and instructions becoming unmanageable as the number of skills grows.
It is installed with pip, initialized with a chosen LLM provider, and is written in Python and released under the MIT license.