MCP Servers Can Ship Their Own Operating Manuals

·https://modelcontextprotocol.io/community/working-groups/skills-over-mcp
articleaimcpagent-skillsskill-distributionprogressive-disclosurecode-mode

fitness.dev Code Mode can serve Rubicon coach skills beside its tools, with discovery and on-demand loading instead of a separate skill install.

The Skills Over MCP Working Group is defining how a Model Context Protocol server can distribute the instructions that explain how to use its tools together. The current proposal, SEP-2640, serves Agent Skills through existing MCP Resources instead of inventing another content format or protocol primitive. It is still an open extension proposal, not a finished standard.

The clever bit is the split between discovery and activation. skills/list exposes lightweight frontmatter, resource URIs, and file digests. The agent loads SKILL.md through resources/read only when the workflow applies, then follows references to scripts, templates, or deeper documentation as needed. That gives clients progressive disclosure without stuffing every manual into the context window.

For fitness.dev Code Mode, the useful shape is one connection that supplies both the coding tools and the Rubicon coaching workflow. The coach skill can change with the service, appear automatically when the server connects, and avoid a second install path. The hard part is trust: the draft treats remote skills as untrusted input, requires visible provenance, gates host-side execution, and binds approval to per-file SHA-256 digests. Distribution gets easier without pretending downloaded instructions are safe.

Key Ideas

  • SEP-2640 is a transport binding for the Agent Skills specification, not a competing skill format.
  • skills/list provides catalog metadata while skills/get retrieves one entry, including skills omitted from large or partial listings.
  • skill://<path>/SKILL.md mirrors the local skill directory model, so supporting files remain individually addressable MCP Resources.
  • Per-file digests let a host detect changed, added, or removed files and revoke content-bound approval when a skill changes.
  • Progressive disclosure keeps names and descriptions available for discovery while loading full instructions and references only on demand.
  • MCP Registry coordination is in progress, but installable bundles containing skills, servers, subagents, and configuration remain outside this working group’s scope.
  • Remote skills create a larger attack surface than remote tool calls because instructions can ask the agent to run code on the host; provenance and explicit approval are part of the proposed contract.

Links

← All discoveries