Git, but for your AI agent skills. 🧠

Tired of messy, fragmented, or conflicting agent capabilities? skill-git brings the power of version control directly to your AI workflows.

Track changes, seamlessly roll back bad edits, and—most importantly—merge overlapping local skills into stronger, unified tools. Works out-of-the-box across any AI Agent platform, including Claude, OpenClaw, Gemini, and Codex.

• 🔀 Combine & Conquer: Stop adding duplicate micro-skills. Automatically scan for local similarities and merge them to make your agent smarter and more robust.
• 📦 Self-Contained & Portable: Each skill folder gets its own independent .git repository.
• 🏷️ Semantic Versioning: Every commit is auto-tagged, keeping your skill history perfectly organized.

Claude (Recommended)

/plugin marketplace add KnowledgeXLab/skill-git
/plugin install skill-git@skill-git

OpenClaw

clawhub install KnowledgeXLab/skill-git

Gemini

gemini extensions install https://github.com/KnowledgeXLab/skill-git

Codex

npx skills add KnowledgeXLab/skill-git -a codex

Then initialize:

/skill-git:init

Supports multiple agents via -a <agent>:

/skill-git:init -a claude
/skill-git:init -a gemini
/skill-git:init -a codex
/skill-git:init -a openclaw

/skill-git:commit

The agent analyzes the diff and recommends a patch or minor bump. You confirm before anything is written.

/skill-git:revert humanizer
/skill-git:revert humanizer v1.0.2

Omitting the version defaults to the previous tag. The operation is atomic — a backup is taken first and restored automatically if anything fails.

/skill-git:check humanizer

Detects internal rule conflicts, contradictory configs, and security issues. Returns a structured report.

/skill-git:scan
/skill-git:scan humanizer code-review

Runs semantic analysis across your skills to find pairs with overlapping rules. Each pair is rated ★★★ / ★★☆ / ★☆☆ and results are cached for merge to pick up.

/skill-git:merge
/skill-git:merge humanizer code-review

Running without arguments picks up the latest scan results. Conflicts are resolved interactively. Nothing is written until you confirm.

~/.claude/skills/
├── humanizer/
│   ├── SKILL.md
│   └── .git/          ← per-skill repo, tagged v1.0.0, v1.0.1 …
├── mcp-builder/
│   ├── SKILL.md
│   └── .git/
└── …

~/.skill-git/
└── config.json        ← registered agents and skill paths

Each skill's .git is fully independent — moving or sharing a skill folder preserves its entire version history.

claude --plugin-dir ./