FastMCP 4.2 and Skills over MCP: What Changes for Your Agent
FastMCP 4.2.0, released November 13, adds a server-side catalog for agent skills. Teams can keep a skill beside the tools it describes and let clients retrieve it from that server, instead of distributing a separate folder to every machine. This helps distribute changes; deciding whether to load and execute those instructions still belongs to the host.
The release implements the experimental Skills over MCP extension, associated with SEP-2640. FastMCP 4.2.0 is a framework version, io.modelcontextprotocol/skills is an extension identifier, and the MCP protocol revision is a separate choice. An upgrade of one does not prove support for the others. This guide checks the published release and specification as of November 17, 2026; it does not claim an end-to-end test in an IDE or chat host.
What is actually delivered?
A skill is instructions plus supporting files. MCP tools describe callable operations; a skill can explain when and how an agent should combine those operations. Serving the instruction files does not itself grant tool permissions or cause the host to execute them.
| Layer | What it provides | What you still need to establish |
|---|---|---|
FastMCP Skill and SkillCatalogProvider | Load authored skills and expose a catalog next to the server’s tools | Which directories and files you intend to publish, and who may read them |
skills/list and skills/get | Extension-aware discovery and a skill manifest | Client and server extension negotiation, including partial implementations |
resources/read | Read instruction and supporting-file contents | File availability, size and digest consistency |
Client.read_skill() and to_directory() | Verify a fetched skill and export it to a local directory | Host-specific local installation, review and execution policy |
The release also exposes skill files as ordinary resources for clients that lack extension-aware discovery. A successful resources/read is therefore useful evidence of file access, but it does not demonstrate support for skills/list, automatic discovery, local installation or execution.
Before adopting 4.2
Check Python first: FastMCP 4.2 requires Python 3.11 or newer. In a separate test environment, pin fastmcp==4.2.0 so the API you review is the API you run. Do not replace a host’s existing MCP SDK environment just to try the feature.
The release’s catalog construction uses these imports and calls:
from fastmcp import FastMCP, Skill
from fastmcp.skills.provider import SkillCatalogProvider
skill = Skill.from_directory("skills/deploy")
mcp = FastMCP("Team Skills")
mcp.add_provider(SkillCatalogProvider([skill]))This excerpt is the registration part of a server, not a complete launch example. skills/deploy must already contain your authored skill; inspect all files before publishing it. Use the server skills documentation for runtime and migration details, checking the installed version if that rolling documentation has advanced.
The 4.1 Skills API is deprecated and scheduled for removal in 4.3. Migrate the old providers in fastmcp.server.providers.skills and helpers in fastmcp.utilities.skills. Skill now names the object you author with; the wire model is MCPSkill. Reusing a 4.1 snippet while assuming a 4.2 import is compatible can conceal migration work until the old alias disappears.
Verify the client capability you need
Start from the extension support matrix and the documentation for the exact host version you deploy. Treat a “partial” entry as a reason to inspect its linked implementation. The matrix is maintained upstream; it is not a guarantee that every version of a listed product implements the same lifecycle.
| Check | Evidence to collect | A result that is insufficient |
|---|---|---|
| Discover | Both peers declare the extension; listing returns the intended skill | The server starts without an exception |
| Retrieve | Read the manifest and every expected file | One instruction file is readable |
| Validate | Confirm byte sizes and SHA-256 digests; mismatches are rejected | The CLI process exits successfully without checking its report |
| Adopt in a host | Verify how that host loads, updates, approves and uses the skill | Export to a local folder succeeds |
Inspector can help separate protocol checks from host behavior. In Inspector 2.15.0, the published CLI is included in @modelcontextprotocol/inspector; do not assume its workspace CLI package is separately published. Skills verification uses --verify with --method skills/list or skills/get. For an ad-hoc stdio target, select --protocol-era modern rather than relying on its legacy default. Consult the fixed CLI source for that version’s invocation and Node requirements.
Interpret the verification report as well as the exit code: 0 may mean verified, or unverifiable dynamic content; 7 reports nonconformance such as size/digest mismatch or unreadable files; 8 means verification was incomplete within the budget; 9 reports unverifiable content when digests are required. These are published CLI semantics, not fresh test results from this article. CLI verification does not establish that a particular host will execute a remote skill.
Integrity and trust are different decisions
A SHA-256 match establishes that retrieved bytes match the declared digest. It cannot establish that the publisher is trustworthy or that following the instructions is appropriate. Review instruction changes, publish only intended files, retain the server’s authorization rules, and let the host preserve user approval and least-privilege tool access. Exporting a skill should not bypass an existing local review step.
Test failure paths before depending on the catalog: a missing file, wrong byte size, wrong digest, changing dynamic content and an unsafe export path. Confirm a clear failure and a clean retry with the corrected skill. Do not describe a protocol-level test as cross-host compatibility.
Where WebMCP fits
If your problem is making actions in a browser-based product available to an agent, explore WebMCP’s English guides and Chinese guides. That browser integration has its own API and lifecycle; the ability to discover remote Skills over MCP does not establish WebMCP support, and a WebMCP implementation does not establish support for this extension.
For the protocol contract, continue with the Skills specification; for operations exposed by a server, use tools. If you also run the official Filesystem, Git or Fetch reference servers, check their separate November security update before testing with real data.