MagicSkills provides a local-first skill infrastructure for multi-agent projects, addressing the common issue of skill duplication and divergence across agents. It transforms scattered SKILL.md files into a reusable, composable, and syncable capability library. This tool is designed for developers building multi-agent systems, offering a unified approach to managing skills that benefits projects requiring consistent capabilities across diverse agent frameworks and applications.

How It Works

MagicSkills operates on a three-layer model: Skill represents a single skill directory containing SKILL.md; Skills is an operable collection of skills curated for a specific agent or workflow; and REGISTRY manages the persistence and retrieval of multiple named Skills collections. This architecture decouples the total installed skill pool from agent-specific subsets, enabling skills to be installed once and reused across various agents and frameworks, whether they integrate via AGENTS.md or direct tool/function calls.

Quick Start & Requirements

• Install: pip install MagicSkills or from source (git clone ..., python -m pip install -e .).
• Prerequisites: Python 3.10-3.13, Git (for installing skills from remote repositories).
• Setup: Local-first installation of skills into a shared pool, creation of named skill collections, and optional syncing to AGENTS.md or exposure via CLI/Python API.
• Links: CLI Docs, Python API Docs.

Highlighted Details

• Enables skill reuse across a wide range of agent frameworks (e.g., AutoGen, CrewAI, LangChain) and IDEs (e.g., Claude Code, Cursor).
• Supports integration via AGENTS.md with none or cli_description sync modes, or direct programmatic integration using skill-tool or Python API.
• Facilitates building a shared skill ecosystem through uploadskill for contribution and install for reuse.

Maintenance & Community

Initiated and maintained by Narwhal-Lab, Peking University. Contribution guidelines are available in CONTRIBUTING.md, with GitHub Discussions encouraged for feature ideas and issues/PR templates provided.

Licensing & Compatibility

Licensed under the MIT License. This permissive license generally allows for commercial use and integration into closed-source projects.

Limitations & Caveats

The execskill command executes in the current process working directory, not automatically within the skill's directory, requiring manual context management if commands are directory-dependent. Skill name conflicts are resolved by using explicit file paths instead of names.