Can you serve skills over MCP to Claude Desktop?

Three ways to get skills into Claude Desktop

From the Claude Help Center, read 5 November 2026. Local servers set up in claude_desktop_config.json aren't available in Cowork or on the web.

How a server serves a skill

MCP servers offer three things to an AI: resources (content), prompts (templates for the user) and tools (functions the AI calls). A skills server typically gives the AI a tool to list or search skills by their descriptions and a tool to fetch one skill's full instructions. Claude reads the list, picks the skill that matches your request, fetches it, and follows it, the same way it uses a skill you uploaded, except the copy lives on the server.

In September 2026 the MCP project finalized a skills extension (SEP-2640) that standardizes how servers publish skills. Not every client implements it yet; the MCP extension support matrix lists which do. A server that also offers the plain tools works with clients that don't.

A worked example: one copy for Desktop, web and a coding agent

Illustration: a made-up business. An eight-person consultancy keeps a "client proposal" skill. Partners use Claude Desktop, analysts use Claude on the web, and one developer uses Claude Code. With uploaded ZIPs, the proposal skill existed in five versions within a month.

1. They host the skills on one MCP server with sign-in, so only staff can connect.
2. The office manager, a Team owner, adds it under Organization settings > Connectors; each person connects once from Customize > Connectors.
3. The developer adds the same server address in Claude Code.
4. A partner asks Claude Desktop for "a proposal for the Hartley renovation"; Claude finds the proposal skill on the server and follows it.

When a partner changes the fee wording, she edits the skill once on the server. The next proposal anyone writes, in any of the three apps, uses the new wording.

Before you connect a server

1. Know who runs it. Anthropic warns that custom connectors reach services it hasn't verified, and a malicious server can try to inject instructions.
2. Check what it asks for at sign-in, and what its tools can change.
3. Prefer read-only tools for skills; editing can happen elsewhere.
4. On Team and Enterprise, let an owner add it once for everyone, rather than each person adding their own.

If you'd rather not run a server, AI Skill Pod is a hosted one: Claude on every plan, ChatGPT on Pro, Business and Enterprise, and coding agents connect to it with one link, from Free (up to 3 members and 10 skills). For GitHub-based sharing instead, see how to share Claude skills on GitHub; for what MCP is, MCP, explained.

Questions

Can I serve skills over MCP to Claude Desktop?

Yes. An MCP server can list your skills and return a skill's instructions when Claude asks for them. Add it as a custom connector in Customize > Connectors, which works in Claude Desktop, the web and Cowork on every plan, or as a local server in Claude Desktop's configuration file, which works in the desktop app only.

Is serving skills over MCP better than uploading them to Claude?

For one person, uploading is simpler. Serving them over MCP pays off when several people, or several AI tools, need the same current version: you edit one copy and every connected AI reads it.

What is the MCP skills extension?

A standard way for MCP servers to publish skills, finalized as SEP-2640 in September 2026. Client support varies; check the MCP extension support matrix. Servers that expose skills through ordinary tools work with clients that don't implement it yet.