Agent Skill Loader

What back1ply/agent-skill-loader MCP server does

The back1ply/agent-skill-loader MCP server connects a collection of Claude Code skills with MCP clients such as Claude Desktop and Cursor. It discovers skill folders containing SKILL.md files, extracts their available descriptions, and exposes the results for agent use. Agents can locate a relevant skill, load its markdown instructions, and optionally copy the skill into the current project.

Skills can be presented as MCP Prompts, which allows clients that support prompts to show them as slash commands. They are also available through programmatic MCP operations. This makes the server suitable for shared skill directories where agents should choose instructions at runtime rather than receive a manually maintained copy.

How it works

At startup, the server combines its standard Claude plugin cache location with paths from configuration. It can also read additional locations from skill-paths.json in the workspace. The effective paths are scanned for skills, and an optional search query can filter the discovered results by name or description.

After finding a skill, an agent can retrieve the contents of its SKILL.md. Installation copies the selected skill folder into .agent/skills/<name>. The destination must remain inside the current workspace, providing a path boundary for this operation. Search paths can be added, removed, or listed while the server is running.

A file watcher monitors configured directories by default. When skills are added or removed, the server sends a list-change notification so supporting clients can refresh their available prompts. Set MCP_NO_WATCH=1 when file watching is unsuitable, such as in a CI environment.

Setup and configuration

Node.js 18 or newer is required. The documented npm setup installs the package globally:

npm install -g agent-skill-loader

Register the agent-skill-loader command in the MCP client's configuration. Alternatively, clone the repository, install its dependencies, and run its build process; the resulting build/index.js can be launched with Node.js.

Configuration accepts MCP_SKILL_PATHS, which can contain a JSON array or a semicolon- or comma-separated list of extra skill directories. MCP_WORKSPACE_ROOT overrides automatic workspace detection. The default search location is %USERPROFILE%\\.claude\\plugins\\cache, and runtime path changes are stored in skill-paths.json.

Tools and capabilities

The README documents discovery, loading, installation, search-path administration, and diagnostics. Discovery supports an optional case-insensitive substring filter. Diagnostic output reports the workspace root, configured paths, whether each path exists and is readable, discovered skill count, and scan warnings.

The supplied listing identifies query and target_path as exposed MCP tools. The README additionally describes operations named list_skills, read_skill, install_skill, manage_search_paths, and debug_info; verify the names exposed by the version you deploy before writing client-side tool calls.

Limitations and notes

The server works with skills represented by folders containing SKILL.md; the provided material does not describe support for other skill formats. Prompt-based use depends on the MCP client supporting MCP Prompts. Installation is restricted to the active workspace, and invalid or unreadable search paths are reported as warnings rather than silently treated as valid sources.

The project is distributed under the MIT license and does not document an API key or external paid service requirement. Its npm package requires Node.js 18 or later. The back1ply/agent-skill-loader MCP server can disable watching, but the material does not describe a separate remote-hosted deployment mode.