The three-tier discipline — search, load, supplement — depends on a master skill file dropped into the agent's project root. Without it, the README says agents skip the score thresholds, load skill bodies speculatively and ignore the tier-3 manifest, which burns exactly the context the design was meant to save. Two more constraints worth planning for: script execution only works against the local Python server, and search quality is only as good as the trigger phrases written into each `SKILL.md` — one vague skill degrades results across the registry.
A registry that serves Agent Skills — the `SKILL.md` format — over MCP instead of copying skill files into every repo and tool. An agent searches the registry by what it is trying to do, reads the confidence scores, and loads the full playbook for the one that matches.
- Seven tools in three tiers — `skills_find_relevant(query, top_k)` for semantic search and `skills_list_all(limit, offset)` for browsing
- `skills_get_body(skill_id, version?)` returns the full instructions in one call, plus a manifest of what else the skill carries; `version` pins to a specific release
- `skills_get_options(skill_id)` returns config schema, variants, dependencies and limitations
- `skills_get_reference(skill_id, filename)` and `skills_get_asset(skill_id, filename)` fetch supporting docs and templates only when the instructions name them
- `skills_run_script(skill_id, filename, input_data)` executes a bundled script — local server only
- 30+ bundled skills distilled from official documentation, each carrying complexity level, time estimate, prerequisites, use cases and a source URL
- Only the frontmatter is embedded — description plus trigger phrases, around 100 tokens — which keeps the search space clean of instruction prose
You deploy your own. It needs a free Qdrant Cloud cluster (URL and API key) and a Cloudflare account — the Worker runs on the free plan, which covers 100k requests/day, because it uses SQLite-backed Durable Objects. Python 3.11+ runs the seeding script and the optional local server; Node.js 18+ provides the `wrangler` CLI. Published as skill-mcp version 1.2.0. Deploy with the one-click button or `bash scripts/setup.sh`, then seed Qdrant with `python -X utf8 -m skill_mcp.seed.seed_skills` using `QDRANT_URL`, `QDRANT_API_KEY`, `WORKERS_AI_ACCOUNT_ID` and `WORKERS_AI_API_TOKEN`. Clients connect over SSE to `https://skill-mcp.<your-subdomain>.workers.dev/sse`; `skills_run_script` needs the local Python server instead, since Workers cannot spawn subprocesses.
One command plus a key — bash scripts/setup.sh, then supply credentials