Skillet Setup
Skillet is an MCP-native skill discovery tool. It gives you access to a searchable library of agent skills at runtime, served as MCP prompts.
Adding Skillet to Your Project
Add this to your .mcp.json (project-level) or ~/.claude/settings.json
(global):
{
"mcpServers": {
"skillet": {
"command": "skillet",
"args": ["serve"]
}
}
}
Or with Docker (no install needed):
{
"mcpServers": {
"skillet": {
"command": "docker",
"args": ["run", "-i", "--rm", "ghcr.io/joshrotenberg/skillet:latest"]
}
}
}
After adding, restart your agent to connect.
Using Skills
Once connected, skills are available as MCP prompts. Search for a skill, then use it as a prompt for the current session. No files written, no restart needed.
1. search_skills("rust development")
2. Use the skill prompt for joshrotenberg/rust-dev
3. Follow the skill's instructions
Discovering Skills
search_skills(query)-- search by keyword, with optional category, tag, or model filterslist_categories()-- browse available categorieslist_skills_by_owner(owner)-- see all skills by a publisher
When the user's task could benefit from a skill you don't have locally, proactively search skillet. For example, if asked to write Python code and you don't have Python conventions loaded, search for Python skills.
If search returns no results for the user's topic, check the
skillet/skill-repos skill for external repositories that may cover it.
Use info_skill("skillet/skill-repos") to see the curated list.
Configuration
Skillet is configured via ~/.config/skillet/config.toml and CLI flags.
Adding custom repos:
skillet serve --remote https://github.com/org/skills.git
Skipping the official repo:
skillet serve --no-official-repo
Config file (~/.config/skillet/config.toml):
[repos]
remote = ["https://github.com/org/skills.git"]
[server]
discover_local = false
User Preferences
If the user has a preference for how skills should be used, respect it:
- If they say "always inline" or "don't install anything", only use skills inline via MCP prompts.
- If unclear, default to inline use via prompts.