Claude Skills MCP: semantic search for skills
Claude Skills MCP
An MCP server with vector search: the agent describes a task and gets matching skills from the Anthropic and K-Dense collections
- 411 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The server downloads packages and skill content from GitHub and the agent reads it as instructions, so connect trusted sources only
- Through read_skill_document the agent receives scripts and data from third-party skills and may run them
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/claude-skills-mcpDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The server downloads packages and skill content from GitHub and the agent reads it as instructions, so connect trusted sources only
- Through read_skill_document the agent receives scripts and data from third-party skills and may run them
This is third-party code. Review the repository files before installing.
What it does
Claude Skills MCP picks skills by the meaning of a task. The agent describes what it wants to do, and the server compares that description with vector embeddings of the skills and returns the matching ones. MCP clients such as Cursor and Codex get access to the Agent Skills format from Anthropic this way. There are three tools. find_helpful_skills searches, read_skill_document fetches individual skill files such as scripts, data and references, and list_skills shows the whole loaded collection. A skill loads progressively: metadata first, then the full text, then its files. By default the server gathers more than 160 skills from the Anthropic repository and the K-Dense scientific collection. GitHub repositories of your own and local folders can serve as sources too. Once an hour the server checks the sources and rebuilds the index if anything changed. No API keys are needed.
Who it is for. For developers and researchers who have collected many skills and need to find the right one by task description in any MCP client.
Good fit when
- Codex or another client without skill support needs the Anthropic skills and wiring each one by hand is tedious
- You need the K-Dense scientific skills for bioinformatics, cheminformatics and data analysis
- You have hundreds of skills across a folder and repositories and want search by task description
- You want the agent to pull a skill's scripts and references itself through read_skill_document
- You need one indexed skill collection for a team, refreshed from GitHub
Not a fit when
- Your agent already reads skills from a folder: Claude Code, Cursor, Windsurf and Copilot do it natively
- You need a maintained project: the repository is no longer developed
- The machine is weak or disk space is tight: the server needs a package with PyTorch and sentence-transformers
- You need the skill to run its scripts: the server finds and serves files, and running them is up to the agent
Example request
Find skills for sequencing data analysis and show which scripts the best match includes.MCP
- Transport
- stdio
- Authentication
- not required
Limitations
The author announced that the project is no longer hosted or maintained: skills now work natively in Cursor, Windsurf, Claude Code and Copilot, so the bridge is not needed. The final version 1.1.0 stays on PyPI, and new issues and pull requests may go unanswered.
Python 3.12 or 3.13 is required. The backend weighs about 250 MB because of PyTorch and the vector search models. The remote backend the author hosted is switched off, so plan on running it locally.
How to disable. Remove the claude-skills entry from the client config, for example mcp.json in Cursor, and restart the client. Downloaded packages and the GitHub cache stay on disk and can be deleted by hand.
FAQ
How do I install Claude Skills MCP in Cursor?
Add the server through the Cursor Directory or put a claude-skills entry in ~/.cursor/mcp.json with the command uvx and the argument claude-skills-mcp. No separate package install is needed, uvx downloads it on its own.
Do I need this server if my agent already reads skills itself?
The author says no. They write that Cursor, Windsurf, Claude Code, Copilot and others support skills out of the box, and they closed the project. The server stays useful when skills have to be searched by meaning in a client without such support.
Which skills are connected by default?
The official Anthropic skills repository, the K-Dense scientific collection and the ~/.claude/skills folder if it exists. Together that is more than 160 skills, and no API keys are needed for them.
How do I add my own skill sources?
Print the default config with the --example-config flag and save it to config.json, add GitHub repositories or local folders to it and start the server with the --config config.json flag.
Why does the server have two packages and why is the first start slow?
The light frontend claude-skills-mcp weighs about 15 MB and answers within five seconds so Cursor does not abort the launch on a timeout. The heavy backend claude-skills-mcp-backend, with vector search, is downloaded in the background by the frontend, which takes roughly one to two minutes.
Discussion
No questions or reviews yet.
Related
MCP serversAn MCP server that opens SKILL.md skills to clients without native support, such as Codex, Copilot and Cursor
A read-only MCP server for the HeyClaude registry: an agent finds skills, MCP servers and hooks and gets ready install commands
A CLI and MCP server that turns doc sites, repos, PDFs and videos into skills and RAG-ready data
A CLI and MCP server for Agent Skills: spec validation, installs from GitHub and on-demand loading of the right skill
A zero-dependency Python CLI: one catalog of MCP servers and skills for every agent, with only a compact index in context
A Claude Code reference plus a collection of over 70 skills for development, security, data and docs, with agent configs
README in short
The repository is a monorepo of two Python packages with detailed documentation: quick start, architecture, tool reference, configuration examples and testing. Separate pages cover the automatic source refresh and workflows with custom skills.