skill-graph-mcp

# Skill Graph **Knowledge graph + vector search for AI agent skill discovery.** Your agent doesn't need to read 500 skill files to find the 3 that matter. Skill Graph indexes `[[wikilink]]`-connected markdown into a traversable graph, then serves it over MCP so agents navigate knowledge structures instead of loading everything into context. ``` search → descriptions → links → sections → full content Most decisions happen before reading a single full file. ``` ## Why AI agents discover skills by dumping every description into the prompt. 50 skills? Fine. 500? You're burning tokens and diluting attention on content the agent will never use. Skill Graph fixes this with **progressive disclosure** — the agent sees only what's relevant, follows links when curious, and loads full content only when it's time to act. In real-world testing, this saves **97-98% of tokens** compared to the "load everything" approach. The graph comes from `[[wikilinks]]` already in your prose. No schema to design, no config to write. If your notes link to each other, you already have a skill graph. ## Install ```bash pip install skill-graph-mcp ``` Or run without installing: ```bash uvx skill-graph-mcp --help ``` ## Quick Start ```bash # Index a directory of markdown files skill-graph index ~/my-skills/ # Search (hybrid vector + graph) skill-graph search "deploy to production" # Browse skill-graph list --limit 20 skill-graph list --category devops # Graph overview skill-graph stats # Start MCP server for agent use skill-graph serve ``` ## Use with AI Agents (MCP) Add to your Claude Code `.mcp.json`, Cursor config, or any MCP-compatible agent: ```json { "mcpServers": { "skill-graph": { "command": "uvx", "args": ["skill-graph-mcp", "serve"] } } } ``` The agent gets 7 tools for progressive disclosure: | Tool | What it does | |------|-------------| | `search_skills(query)` | Semantic + graph search — returns descriptions only | | `get_skill(name)` | Sha…