Serve Agent Skills over MCP with centralized hosting and token-efficient progressive disclosure.
Agent Skills typically live as files in repositories (.claude/skills/, plugins, and so on). This server lets you:
- Host skills centrally: serve skills over HTTP to any MCP client, from local directories, git repositories, or OCI registries
- Minimize token usage: three-tier progressive disclosure loads only what agents need, when they need it
- Isolate sessions: each client connection keeps its own expansion state
export SKILLS_MCP_PATHS="/path/to/skills"
uv run skills-mcpConnect any Streamable HTTP MCP client (Claude Code, Claude Desktop, Roo Code, Cline, Continue) to http://localhost:8080/mcp. See the quickstart for installation options (including Docker) and connect your client for per-client setup.
The same skills are exposed through three complementary MCP surfaces, so a client can use whichever mechanism it supports: resources (skills:// URIs with progressive disclosure), tools (list_skills, get_skill, get_skill_resource, validate_skill), and prompts (one per skill, which clients turn into slash commands). The list_skills tool embeds the skill catalog in its always-loaded description, which is what makes agents pick up served skills on natural prompts without being told.
Details: MCP surface reference, getting agents to use your skills, and the SEP-2640 alignment notes.
A skills.yaml file can pull skills from three kinds of source, in any combination (local takes precedence over git, then OCI, on name collisions):
- Local filesystem (
local:): directories scanned forSKILL.md - Git repositories (
git:): cloned over HTTPS fromgit://host/owner/repo[@ref][#subdir]references - OCI registries (
oci:): skill artifacts pulled from a registry
See skill sources for walkthroughs and the configuration reference for the full schema.
Full documentation lives at stacklok.github.io/skills-mcp (source under docs/):
git clone https://github.com/stacklok/skills-mcp.git
cd skills-mcp
uv sync --all-extras
uv run pytest
uv run ruff check .
uv run mypy src/Proprietary. © Stacklok, Inc. All rights reserved.