Serve Glean skills over the MCP skills extension

[aditya-scio]

Implements OpenAI's Import skills from the MCP server flow, so hosts that import skills from the MCP server (ChatGPT plugins) pick up the skills the signed-in user has in Glean instead of only the one skill packaged under plugins/glean/skills/.

Skills come from the experimental Platform Skills API: GET /api/skills for metadata (cursor-paginated), GET /api/skills/{id}/content for the bundle — same origin and bearer token setup already captured, with X-Glean-Include-Experimental: true.

Surface

capabilities.extensions["io.modelcontextprotocol/skills"] is advertised at initialize, alongside resources (required to register resources/read).

Method	Behavior
skills/list	One page (10 skills); cursor maps to the Platform API cursor. Each entry: skill://glean/<name>/SKILL.md uri, full parsed SKILL.md frontmatter, every bundle file with a sha256: digest.
skills/get	The catalog entry for one SKILL.md uri.
resources/read	Bytes for any listed uri — text for UTF-8, blob otherwise. Served from an in-process cache, with a bounded crawl as fallback when the host reads a uri this process never listed.
resources/list	Only what was already listed, so host startup never triggers bundle downloads.

Unconfigured or signed-out sessions serve an empty catalog; setup remains the only path that drives OAuth. Non-ENABLED, oversized (spec import limits: 256 KiB SKILL.md, 1 MiB/file, 5 MiB/skill, 100 files) and duplicate-named skills are skipped and logged to glean-server.log.

Changes

• src/skills-catalog.ts — catalog build, caching, and the three handlers
• src/zip.ts — minimal central-directory ZIP reader (stored + deflate), so bundles need no new dependency
• src/index.ts — capabilities + registration
• zod promoted to a direct dependency, imported as zod/v4 to match the SDK's entrypoint (bundle grows ~25 KB, not ~400 KB)
• README section

Verification

• npm run typecheck clean; npm test 201 passing (12 new in tests/skills-catalog.test.ts, covering the zip reader, bundle normalization, digests, cursor passthrough, auth headers, skip/crawl paths)
• dist rebuilt and in sync
• End-to-end run of the built server over stdio against a fake Platform API: skills/list → skills/get → resources/read, wrapper directory stripped, advertised digests matching the served bytes

Notes

• The Platform Skills API is experimental; response shape changes would need a matching update here.
• Not implemented: skills list-changed notifications, zip64/CRC verification.

🤖 Generated with Claude Code