MCP Skills Extension

dcc-mcp serves its skills through two parallel discovery tracks over the 2026-07-28 stateless protocol:

Track	Methods	Audience
Tools (existing)	search_skills, list_skills, get_skill_info	every MCP client
Skills extension (new)	skills/list, skills/get, resources/directory/read	hosts implementing io.modelcontextprotocol/skills

Both tracks read the same catalog, so a skill has the same name and the same content on either path. The tools track is unchanged — adopting the extension requires no migration.

The extension is specified against base protocol revision 2026-07-28 or later; legacy 2025-x sessions are unaffected.

Specification

• Skills over MCP overview
• ext-skills specification (stable)
• ADR-010 — dual-protocol migration

Server configuration

The extension is enabled by default on the stateless path. Opt out with McpHttpConfig.enable_skills_extension = false:

from dcc_mcp_core import McpHttpConfig, create_skill_server

cfg = McpHttpConfig(port=8765)
cfg.enable_skills_extension = False
server = create_skill_server("maya", cfg)
server.start()

Because the extension serves its content through resources/read, declaring it also requires the resources capability. Setting enable_resources = false therefore suppresses the extension too, and the three methods answer Method not found.

Capability negotiation

server/discover declares both:

{
  "capabilities": {
    "resources": { "subscribe": false, "listChanged": false },
    "extensions": {
      "io.modelcontextprotocol/skills": { "directoryRead": true }
    }
  }
}

directoryRead: true means resources/directory/read is implemented for every directory inside a served skill.

Resource URIs

Each file of a skill is an individually addressable resource:

skill://<skill-name>/SKILL.md           the skill's entry point
skill://<skill-name>/<relative-path>    any supporting file

The first path segment is the skill's name as declared in its SKILL.md frontmatter, so the name is recoverable from the URI alone. The final segment is the file being addressed (SKILL.md for the entry point, otherwise the supporting file's own name), so it is not the skill name in general. A skill's root directory is the entry-point URI with the /SKILL.md suffix removed and no trailing slash — skill://maya-geo.

skills/list

Returns one complete entry per skill: its SKILL.md URI, the frontmatter verbatim, and a complete manifest of every file with its SHA-256 digest and byte size.

{
  "jsonrpc": "2.0",
  "id": 4,
  "result": {
    "resultType": "complete",
    "skills": [
      {
        "uri": "skill://maya-geo/SKILL.md",
        "frontmatter": {
          "name": "maya-geo",
          "description": "Polygon modelling helpers for Maya",
          "license": "MIT"
        },
        "resources": [
          {
            "uri": "skill://maya-geo/SKILL.md",
            "digest": "sha256:b95a384300adeea2d902f7d19cd7c04b378ef58e09759107b9c7db4dcacbaa25",
            "size": 190
          }
        ]
      }
    ],
    "ttlMs": 0,
    "cacheScope": "private"
  }
}

Notes:

• resources is complete — it lists SKILL.md and every supporting file, each exactly once. A host verifies each file it reads against that entry.
• frontmatter passes every field the skill author wrote through unchanged, including keys dcc-mcp itself does not model.
• Results are paginated with cursor / nextCursor. An entry is atomic: a skill's resources set is never split across pages.

skills/get

Returns the entry for a single skill, named by the URI of its SKILL.md. It does not depend on a prior skills/list, and it answers for any skill the server serves, including one absent from a partial listing.

{ "jsonrpc": "2.0", "id": 5, "method": "skills/get", "params": { "uri": "skill://maya-geo/SKILL.md" } }

An unknown URI returns -32602 (Invalid params), the same code resources/read uses for an unknown resource.

resources/read

Skill files are read with the standard method; no skill-specific read semantics are defined. The bytes returned are exactly the bytes the entry's digest and size describe. SKILL.md is served as text/markdown with text payloads; non-UTF-8 files come back as a base64 blob.

Relative references inside a skill resolve against that skill's root, so references/GUIDE.md in skill://maya-geo/SKILL.md is skill://maya-geo/references/GUIDE.md.

Reading a SKILL.md does not by itself activate the skill — activation is the host's own decision, taken after verifying content and obtaining any approval it requires.

resources/directory/read

Lists the direct children of a skill directory. Non-recursive: descend by calling the method again on a child directory. Subdirectories come back as directory resources with mimeType: "inode/directory".

{ "jsonrpc": "2.0", "id": 7, "method": "resources/directory/read", "params": { "uri": "skill://maya-geo/templates" } }

A URI that does not exist, or exists but is not a directory, returns -32602.

Limits

Limit	Value
Files per skill	512
Total size per skill	16 MiB (16,777,216 bytes)

Both are the values every conforming host must accept. A skill on disk that exceeds either limit is not published through the extension — it stays available on the tools track, and a warning is logged at WARN level rather than served with a manifest a host would reject.

Errors

Code	Cause
-32602	Unknown skill URI, unknown skill file, or a directory read on a non-directory
-32601	Method not found — extension disabled, or resources switched off

Diagnostics

skills/list computes digests from the bytes on disk on every call, so a cache entry can never go stale relative to what resources/read will return. The trade-off is that listing walks every served skill; hosts should honour the ttlMs hint and re-list only when it expires.