From nothing to a working skills server

The quickest path is Docker — nothing to clone or build. All state lives in ./data, mounted into the container, so your skills and config stay on the host.

docker run -d -p 3000:3000 -v ./data:/data \
  -e MCP_SKILLS_TOKEN=your-secret-token \
  vantreeseba/mcp-skills-manager:latest

Prefer to run from source? With Node ≥ 22.18:

git clone https://github.com/cubicecho/mcp-skills-manager && cd mcp-skills-manager
npm install
npm run build
MCP_SKILLS_TOKEN=your-secret-token npm start

On first run, if you don't set MCP_SKILLS_TOKEN, a random bearer token is generated into data/config/settings.json and printed to the logs. The server also seeds a starter getting-started skill so there's something to see right away.

Visit http://localhost:3000 and paste your token when prompted. You'll land on the Skills page — every skill served by the root /mcp endpoint, each of which is both a callable tool and a skill:// resource. The connect card at the bottom shows the exact URL to hand a client.

localhost:3000

Click New skill. Give it a title (which becomes its slug) and a one-line description that tells an agent when to use it. Pick a layout: a single <name>.md file, or a <name>/SKILL.md directory that can hold supporting files. The second tab lets you upload an existing .md, folder, or .zip instead.

New skill

Creating a skill drops you into the editor. Write markdown on the left and watch the rendered preview on the right (or switch to full Edit / Preview). Set the description and tags, and use the Serve on root /mcp toggle to decide whether every client sees this skill or only the workspaces that list it. For directory skills, the file tree above lets you add, upload, rename, and export supporting files.

localhost:3000/skills/code-review

A workspace is a named subset of skills served at its own /mcp/w/<slug> endpoint — so you can hand each agent exactly the skills it needs. Open the Workspaces page, click New workspace, name it, and check the skills to include. You can override the tool-exposure mode per workspace, or disable it to make its endpoint return 404.

New workspace

localhost:3000/workspaces

Point an MCP client at your server. For Claude Code over streamable HTTP:

# all skills
claude mcp add --transport http skills http://localhost:3000/mcp \
  --header "Authorization: Bearer <token>"

# just one workspace
claude mcp add --transport http backend http://localhost:3000/mcp/w/backend \
  --header "Authorization: Bearer <token>"

Or run it locally over stdio with the packaged binary — scope it to a workspace with --workspace:

{
  "mcpServers": {
    "skills": {
      "command": "mcp-skills-stdio",
      "args": ["--data-dir", "/path/to/data", "--workspace", "backend"]
    }
  }
}

Your agent now sees each skill as a tool (call it to load the body) and as a skill:// resource. It can also discover skills first with the list_skills and search_skills meta-tools, then load only what's relevant.

The Settings page shows server status and the main switches: MCP tool exposure (one tool per skill, or a single load_skill loader), agent authoring (let agents create and refine skills over MCP), and HTTP live updates (push resource-change notifications over SSE). Everything here is also editable directly in settings.json.

localhost:3000/settings

That's it — you're serving skills.

Add more skills, split them across workspaces, and let agents author their own. For the full reference — frontmatter, the REST API, resources, and security notes — read the docs.