skillmcp
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
skillmcp
An MCP stdio server that makes skill packs callable by any agent — search them, fetch one, done. No copying folders into a prompt, no giant context dump.
Ships with 50 validated skills about running such a server: wiring clients, designing tools, invoking them cheaply, verifying the protocol, governing side effects and publishing packs.
skillmcp serve --packs-root /srv/packs # JSON-RPC 2.0 over stdin/stdout
skillmcp tools # what this server advertises
skillmcp index --packs-root . # packs and skills it can see
skillmcp call search_skills --args '{"query":"stdout purity"}'Why a server instead of a directory
A skill directory only helps an agent that already knows the folder exists and can afford to read it. A server turns that directory into three calls:
Tool | Returns | Bounded by |
| every pack with category and skill count | number of packs |
| ranked slugs with title, tags and description |
|
| one full | one document |
So the agent searches with keywords, reads one procedure, and acts. Search
covers slug, title, tags and the per-skill description, which is why a query
like stdout purity finds exactly the right skill.
Related MCP server: skills.sh MCP Server
Protocol
Protocol revision pinned to 2025-03-26, newline-delimited JSON-RPC 2.0 over
stdio. Implemented: initialize, notifications/initialized, ping,
tools/list, tools/call. Everything else returns -32601.
$ printf '%s\n' \
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26"}}' \
'{"jsonrpc":"2.0","method":"notifications/initialized"}' \
'{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
'{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"search_skills","arguments":{"query":"stdout purity","limit":2}}}' \
| skillmcp serve
{"id":1,"protocolVersion":"2025-03-26","capabilities":{"tools":{"listChanged":false}},"serverInfo":{"name":"skillmcp","version":"1.0.0"}}
{"id":2,"tools":["list_packs","search_skills","get_skill"]}
{"id":3,"matches":["stdout-purity-check"],"total":1}Notifications never get a reply. Malformed lines get -32700 and the server
keeps serving. stdout carries protocol messages only; diagnostics go to stderr,
which the test suite asserts.
Client config
{
"mcpServers": {
"skills": {
"command": "python",
"args": ["-m", "skillmcp", "serve", "--packs-root", "/srv/packs"]
}
}
}Point --packs-root at a directory of packs, or at a single pack. Anything with
a pack.json is discovered, including packs built by the other repos in this
series.
The 50 skills
Phase | Skills | Focus |
| 8 | Client config, transport choice, version handshake, capability declaration, trust, boundaries, secrets, reload semantics |
| 8 | Tool naming, schema strictness, bounded output, error shape, idempotency, surface minimisation, destructive gating, descriptions as contracts |
| 9 | Search-then-fetch, call budgets, parallel safety, argument exactness, error classification, pagination, determinism, context management, verification |
| 8 | Protocol conformance, lifecycle, schema validation, error paths, transcript fixtures, stdout purity, concurrency, replay |
| 9 | Scope guard, rate limits, credential isolation, audit log, destructive confirmation, retention, least privilege, sandboxing, escalation |
| 8 | Frontmatter contract, one procedure per skill, use/avoid clauses, signals and pitfalls, done-when, references, versioning, local publishing |
skillmcp list --phase govern
skillmcp search stdout
skillmcp show protocol-conformance-testEach skill renders deterministically from a Python corpus dict, carries a
SHA-256 in pack.json, and is re-validated against six floors on every run.
Install
python -m venv .venv && . .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest -q
skillmcp serve --packs-root .Python 3.10+. No runtime dependencies, and the server never opens a socket — stdio only.
Safety
The server is read-only: three tools, no writes, no network. It reads files
under the packs root you give it and returns text. Credentials never belong in
the client config; a server that needs one should read it from the environment,
as secrets-via-environment describes.
MIT licensed.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Related MCP Connectors
Search and load AI agent skills. Free metadata; full content requires a Pro or Studio account.
Search and fetch curated AI agent skills, prompts and MCP servers.
Executable skill packages for agents: vetted competence, ready to run, not just documents.
Agent skills for Claude, Codex and Cursor: find, install and load skills from one connector.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables coding agents to search, recommend, and validate a stack of reusable skills from a local catalog, producing deterministic plans without modifying the project.1MIT
- AlicenseAqualityCmaintenanceEnables AI agents to autonomously search, evaluate, and install skills from the skills.sh catalog.413 npmISC
- AlicenseNot gradedqualityBmaintenanceEnables MCP clients to discover, search, and load agent skills from many sources, including other agents' skill directories, Claude Code plugin marketplaces, and remote git repositories, with no copying or syncing required.806 npm7MIT
- AlicenseAqualityCmaintenanceEnables LLM agents to discover and search a catalog of skills, then fetch skill instructions and supporting references, scripts, and assets on demand.10MIT