Agent Skills MCP

A remote MCP server that hands agent skills to Claude on request.

Ask "I need a skill for local lead research" and Claude calls get_skill, reads the instructions, asks you for what the skill needs, and gets to work.

Skills live in a database rather than in the deployment, so editing one is a seed command, not a redeploy.

What it does and does not do

This server serves skills. It does not run them, and that is deliberate.

A skill needs the user's own API tokens and writes files to the user's own machine. A remote server has neither. So the split is:

The server sends instructions. The agent on the other end carries them out with credentials this server never sees.

Tools

Setup

1. Create the table. In your Supabase project, open the SQL editor and run supabase.sql.

2. Seed it from the .SKILL files in skills/, which are the source of truth:

cp .env.example .env && npm install && npm run seed

3. Deploy to Vercel. Import the repo, then set two environment variables:

The service role key stays on your machine for seeding. The server only reads, so it only needs the anon key.

4. Connect it to Claude. Settings → Connectors → Add custom connector, and paste your endpoint:

https://your-app.vercel.app/api/mcp

Editing a skill

Edit the file in skills/, then:

npm run seed

Live within a minute, no redeploy. Responses are cached for 60 seconds so a burst of calls hits the database once.

Adding a skill

A skill is markdown with YAML frontmatter, saved with a .SKILL extension:

---
name: my-skill
description: One sentence on what it does and when to use it. Agents read this to decide whether the skill is relevant, so it matters more than it looks.
---

# My skill

Instructions for the agent...

Drop it into skills/, run the seed, done.

Local development

npm install
npm run dev

The endpoint is http://localhost:3000/api/mcp. Point the MCP Inspector at it to try the tools without a Claude connector:

npx @modelcontextprotocol/inspector

Security

• No secrets pass through the conversation. The server holds no user credentials and has no tool that accepts one.

• The database is read-only in production. Row level security allows anonymous select and nothing else; writes need the service role key, which never reaches the deployment.

• The endpoint is unauthenticated. Skills are public instructions, so anyone with the URL can read them. Put auth in front of it before storing anything you would not publish.

License

MIT