A slash command injects the skill directly; the tools exist for when the agent has to find one it did not know about. debug_info is the tool that earns its place in practice — when nothing is discovered, it tells you which directory it looked in and why the read failed, instead of leaving you to guess.
A bridge between a folder of skill files and an MCP client. Skills are discovered from the directories you configure, published both as prompts and as tools, and re-announced automatically when you add or remove one.
- Skills published as MCP prompts, so they appear as slash commands and inject their content with no tool call at all
- list_skills — the available skills, with an optional query filtering by name or description
- read_skill — the markdown instructions for one skill
- install_skill — copy a skill folder into the current project, restricted to paths inside the workspace
- manage_search_paths — add, remove or list the directories scanned, at runtime, without editing a config file
- debug_info — the search paths in use, whether each exists and is readable, how many skills were found, and the warnings raised while scanning
- A file watcher that fires a change notification when skills are added or removed, so the client's list stays current
Node.js 18 or higher. Install globally from npm as agent-skill-loader and register the command, or build from source and point at the built entry file. It detects its workspace and scans the default plugin cache path plus anything you add; MCP_SKILL_PATHS adds more paths, MCP_WORKSPACE_ROOT overrides the detected root, and MCP_NO_WATCH turns the watcher off for CI.
One command — npm install -g agent-skill-loader