The metahub MCP server plugs into your AI client (Claude Code, Cursor, Claude Desktop, Continue, anything that speaks MCP). Once it is in, your AI can find, fetch, and wire up skills, MCP servers, agents, and plugins on its own.
claude mcp add metahub -- npx -y @metahub/mcp-serverThat snippet is for Claude Code. Need Cursor, Claude Desktop, or Continue? See below.
One command for terminal-first clients. One JSON block for editor-first clients. Restart the client and you're done.
claude mcp add metahub -- npx -y @metahub/mcp-serverRun once. Claude Code picks it up on next start.
Any MCP-aware AI client works. If yours isn't listed, use the Generic stdio tab and point your client at npx -y @metahub/mcp-server.
Each tool maps to something natural to say. You don't memorize tool names; the AI picks the right one.
metahub_search“find me a skill for parsing PDFs”
Searches the public catalog by intent. Filters by kind, ranks by usage and rating.
metahub_install“install that”
Generates the install command (or, with consent, runs it) for the artifact you just found.
metahub_list_installed“what do I have installed?”
Reads ~/.metahub/installs.json locally and answers from your machine.
metahub_signin“sign in”
Walks you through the GitHub Device Flow to unlock review and publisher tools.
metahub_submit_review“leave a 5-star review on the pdf skill”
Posts a verified review back to metahub. Requires sign-in.
metahub_my_statsPublisher“how is my pdf-parser doing?”
Pulls live observability for your published artifacts. Publisher only.
You don't sign in to browse, search, or install. The MCP server only asks for credentials the moment you cross into territory that has to be tied to a real human (reviewing, publishing, viewing your private dashboard).
Every search, every install, every catalog read. Works the moment the MCP server starts.
Asking to review or to see your own observability triggers a one-time sign-in via the AI itself ('sign in').
The GitHub session token writes to ~/.metahub/config.json. No restart needed; subsequent tools pick it up.
Stdio only for publisher tools
Reviews, publisher stats, and the “what have I published” query run only on the local stdio MCP server because they read your local token. A hosted HTTP transport can't see your machine.
The MCP server is a read client. It does not collect invocation telemetry the way the artifact SDKs do, because here you are the one running it.
Want to point the server at a private mirror? Set METAHUB_REGISTRY_URL. The full env-var reference is in the MCP server docs.
No. The MCP server installs artifacts directly using the same library mh does (@metahub/installer). The CLI is a separate optional path for terminal users; you never need it just to install artifacts via MCP. The CLI is also available if you'd rather drive installs by hand.
No. Search and install are anonymous. You only sign in if you want to leave reviews or see your own publisher observability, and the AI walks you through it when you ask.
Any MCP-aware client. We document Claude Code, Cursor, Claude Desktop, and Continue explicitly. For everything else, the Generic stdio tab above works the same way. See the full client list.
Your install list and login token sit at ~/.metahub/ (or %USERPROFILE%\.metahub\ on Windows) on your machine. The MCP server reads from there; it never uploads them. The catalog data is hosted and public.
In Claude Code: claude mcp remove metahub. In Cursor, Claude Desktop, or Continue: delete the metahub entry from the client's MCP config file and restart the client.
Paste this into Claude Code, then ask your AI to find a skill. That's the whole flow.
claude mcp add metahub -- npx -y @metahub/mcp-server