Final release of Claude Skills MCP. This project is no longer hosted or maintained — Agent Skills are now natively supported by all major AI platforms — but v1.1.0 remains installable from PyPI.
Backend
- Tarball-based GitHub loading: repositories are downloaded as a single codeload tarball (not counted against the GitHub API rate limit) and extracted to a 24h snapshot cache. Replaces per-file raw fetches + tree API calls; skill documents are read lazily from the local snapshot. Warm starts make zero network requests.
- Auto-update fix: the hourly updater now bypasses the snapshot cache (
force_refresh), so detected changes actually reload. - YAML frontmatter parsing:
parse_skill_mdnow usesyaml.safe_load(with regex fallback), fixing multilinedescription: >/|scalars that previously indexed as a literal">"and made those skills unfindable in search. - Embedding disk cache: skill embeddings persist per (model, text batch); warm starts skip encoding and model loading entirely.
- Search quality: embeddings now use
name: description; cosine similarity guards against zero-norm vectors.
Frontend
- Dynamic port selection: the spawned backend gets a free ephemeral port by default (pin with
--port); all kill-by-port logic removed — cleanup only ever targets the spawned process. Multiple MCP clients can now run the server side by side. - Faster readiness: the proxy connects as soon as the backend HTTP server is up instead of waiting for all skills to load; the backend reports loading progress to tool calls.
--remote <url>implemented: connect to an already-running backend (with retry) instead of spawning one locally.
Project
- Python 3.12–3.13 supported (
requires-python >=3.12,<3.14). - Deterministic offline test suite: live-GitHub tests are behind a new
networkpytest marker (deselected by default); tarball loading, YAML parsing, and frontend/backend tool-schema sync are covered by new no-network tests. - CI runs offline integration tests strictly; live-GitHub tests are advisory.
Full Changelog: v1.0.6...v1.1.0