How to Load Agent Skills from an MCP Server (SEP-2640)

You can load agent skills from an MCP server today only if your client supports MCP's Skills extension, defined in SEP-2640, and as of Oct 30, 2026 that is still a short list. The MCP project's support matrix shows partial Skills support in ChatGPT, fast-agent and MCP Inspector. The ext-skills implementations list adds a few more developer tools, such as MCPJam Inspector and the mcpc CLI. The matrix lists no Skills support for Claude Code, Claude Desktop, Cursor or VS Code GitHub Copilot, and Codex is not on it at all, so in those tools skills still come from folders on disk or from plugins. The MCP project's own page says SDK and host support is still being implemented, so expect this list to change. This guide is for people who use skills, not people who build servers: what the extension does from your side, the steps where it works, and what to do where it doesn't.

Based on the MCP Skills extension docs, SEP-2640, the ext-skills implementations list, the MCP extension support matrix, and the fast-agent, Claude Code, Cursor and Codex docs, as of Oct 30, 2026.

Why answers to this question disagree

Search for "agent skills MCP server" and you get three different things under the same words.

1. The official Skills extension. The server lists its skills, and a supporting client fetches a skill's SKILL.md and files when a task needs them. This is SEP-2640, extension ID io.modelcontextprotocol/skills. It needs support in your client.

2. Skills served as MCP prompts. Some servers expose each skill as an MCP prompt. Any client that shows MCP prompts can run them, but a prompt is a single block of text, without a skill's folder of scripts and reference files.

3. Skill search servers. Some directories and open-source projects run an MCP server with their own tools, with names like search_skills or get_skill. These work in any MCP client because they are ordinary tools, but each one defines its own behavior.

The rest of this guide is about the first kind, because it is the shared standard. The last section covers the other two as workarounds.

What the Skills extension does, from your side

A skill served over MCP is the same thing as a skill on disk: a folder with a SKILL.md file whose frontmatter gives a name and a description, plus any scripts or reference files it uses. The extension only changes where the folder comes from.

When your client connects to a server that supports the extension, it can ask for a list of the server's skills, each with its frontmatter, an address such as skill://git-workflow/SKILL.md, and the size and SHA-256 fingerprint of every file. The folder name just before SKILL.md, git-workflow in this example, is always the skill's name. When a task matches a skill, or when you pick one yourself, the client fetches that skill and reads only the files it needs. Clients must not download a skill's files ahead of need: not when they connect, not when they list skills, and not when you approve one.

Your client decides how skills show up. The extension leaves the user experience to the host: the model can pick a skill by its name and description, or you can pick it. Reading a skill's SKILL.md does not by itself turn the skill on. The host loads it through its own skill-loading path, checks each file against the listed size and fingerprint, and applies any approval it asks you for. For a skill nested inside another skill, hosts must get your consent again before turning it on.

The SEP behind this was created April 23, 2026, on behalf of the Skills Over MCP Working Group. It has Final status; the working group's charter says it was merged on September 13, 2026.

How to load skills from an MCP server: the steps

1. Check that your client supports the extension. Look for your client on the MCP extension support matrix and the ext-skills implementations list (both are linked below). If it is not there, skip to the last section.

2. Connect a server that serves skills. Add it the way your client adds any MCP server. The server has to support the extension too. On the implementations list, Hugging Face's hf-mcp-server is one that implements the released version, over HTTP only.

3. Browse the skills the server lists. How this looks depends on the client. In fast-agent, for example, /skills registry lists the skills a connected server offers.

4. Install or use a skill. In fast-agent, /skills add installs a listed skill: it downloads every file the server lists for it and writes a local copy only if each file matches the byte size and SHA-256 digest the server listed. A SHA-256 digest is a short fingerprint of a file's contents, so a match means the file arrived exactly as the server described it. It does not prove the publisher is trustworthy or the skill is safe. Other clients may load a skill straight into the session instead of saving a copy.

5. Read the skill before you use it. A skill is a set of instructions for your agent, so open its SKILL.md and any scripts it includes to see what it will ask the agent to do, the same as you would for a skill you copied from a repository.

Which clients support skills over MCP (Oct 30, 2026)

From the MCP extension support matrix and the ext-skills implementations list. Both are maintained by the community, so treat them as a snapshot.

Partial support on the matrix: ChatGPT, fast-agent and MCP Inspector. The implementations list rates fast-agent as implementing the released version, and MCP Inspector as partial since version 2.6.0. For ChatGPT it notes that plugins import a server's skills once, when the plugin is submitted, as a fixed copy.

Released version on the implementations list, but no Skills entry on the matrix: MCPJam Inspector, plus smaller tools such as the mcpc CLI.

In progress (an open pull request): Goose, opencode, Agno and Cherry Studio, among others.

Planned: VS Code, where support has been requested for GitHub Copilot but has no pull request yet.

Listed on the matrix with the Skills column blank (no support listed): Claude on the web, Claude Desktop, Claude Code, VS Code GitHub Copilot, Microsoft 365 Copilot, Goose and Cursor, among others.

Not on the matrix or the implementations list: Codex.

On the server and SDK side, the TypeScript SDK ships the extension in version 2.4.0. Support is merged but not yet released in the Python and C# SDKs, and still an open pull request in the Go, Rust, Java, PHP and Ruby SDKs. FastMCP's SkillsProvider is available as an experimental, opt-in feature. That matters to you because many servers are built on those SDKs, so support should spread as they ship.

If your client doesn't support it yet: Claude Code, Cursor and Codex

Install the skill as a folder. Claude Code, Cursor and Codex all read skills in the SKILL.md format from skill folders on disk. If a server's skill also lives in a public repository, copy or install the folder once, and the agent loads it when a task matches.

Use MCP prompts for short, repeatable instructions. Claude Code lists each MCP prompt from a connected server as a command, shown as /servername:promptname (MCP), and you can also run it as /mcp__servername__promptname. This is how servers that serve skills as prompts reach Claude Code today. One catch from the Claude Code docs: prompts from a server named anthropic-skills do not appear, because Claude Code reserves that name for skills synced from claude.ai. The server's tools still work.

Use a skill search server as an ordinary MCP server. Servers that offer their own tools for finding skills work in any MCP client, Claude Code and Cursor included, because the client only sees normal tools. Read each server's docs, since there is no shared standard for what those tools do.

In Codex, pair a skill with a server. A skill can declare the MCP server it depends on in its agents/openai.yaml file, and a plugin can bundle skills together with MCP server settings. The skill itself still lives in the folder or plugin.

FAQ

Can an MCP server serve agent skills to Claude Code or Cursor?

Not through the official Skills extension as of Oct 30, 2026. Claude Code's docs do not mention it, and the MCP matrix lists no Skills support for Claude Code or Cursor. Both can run MCP prompts and tools from a server, and both load skills from SKILL.md folders.

What is SEP-2640?

It is the MCP proposal that defines the Skills extension, io.modelcontextprotocol/skills. It has Final status. Extensions are optional, so a client or server supports it only if it declares it in its capabilities.

Do skills over MCP use a different SKILL.md format?

No. The extension serves the same SKILL.md folders, following the Agent Skills specification, with name and description in the frontmatter, so a skill can work from disk or from a server.

What is a skill:// URI?

It is the address an MCP server gives a skill, such as skill://git-workflow/SKILL.md. The folder name just before SKILL.md is the skill's name. Clients must not look the first part up as a web address.

Is a skill served as an MCP prompt the same as a skill?

No. A prompt is one block of text you run as a command. A skill is a folder that can include scripts and reference files, and the agent can pick it when a task matches.

Which MCP servers serve skills today?

As of Oct 30, 2026, the ext-skills implementations list shows more than a dozen servers on the released version, including Hugging Face's hf-mcp-server, the Cloudflare API MCP server, Outline and n8n.

Let your agent search Skill Store 33

Sources

MCP Skills extension overview: https://modelcontextprotocol.io/extensions/skills/overview

SEP-2640 Skills Extension: https://modelcontextprotocol.io/seps/2640-skills-extension

MCP extension support matrix: https://modelcontextprotocol.io/extensions/client-matrix

ext-skills implementations list: https://github.com/modelcontextprotocol/ext-skills/blob/main/docs/implementations.md

Skills Over MCP Working Group: https://modelcontextprotocol.io/community/working-groups/skills-over-mcp

fast-agent, Skills over MCP: https://github.com/evalstate/fast-agent/blob/main/docs/docs/mcp/skills-over-mcp.md

Claude Code, MCP: https://code.claude.com/docs/en/mcp

Claude Code, skills: https://code.claude.com/docs/en/skills

Cursor, skills: https://cursor.com/docs/skills

Codex, build skills: https://learn.chatgpt.com/docs/build-skills

Agent Skills specification: https://agentskills.io/specification

This article was created with the help of AI (AI-Assisted). Every config key and command was checked against the official docs in the Sources list. ReadyAI is accountable for the content.

Keep reading