# Pryno MCP

Pryno MCP is a remote [Model Context Protocol](https://modelcontextprotocol.io) server. It lets any MCP client search the Pryno directory of public MCP servers and score a website for AI-agent readiness.

The directory is filled by a separate private crawler. This server only reads that database. It does not crawl, and it does not store CRM credentials.

- Remote endpoint: `https://mcp.pryno.ai/mcp`
- Source: [github.com/Pryno-ai/pryno-mcp](https://github.com/Pryno-ai/pryno-mcp)
- Website: [pryno.ai](https://www.pryno.ai)

## Tools

| Tool | What it does |
| --- | --- |
| `search_mcp_servers` | Search by text, category, transport, language, official servers, or remote servers. |
| `get_mcp_server` | Full record for one slug, including packages and stdio install hints. |
| `list_categories` | Categories and how many servers are in each. |
| `directory_stats` | Total servers, servers added in the last 7 days, and the last crawl time. |
| `run_readiness_check` | Score a public URL from 0 to 100. |

Every tool is read-only. Search and lookup results include `readiness_score` and `last_checked_at` when the crawler has stored them.

Example: search for Postgres, then open one result.

```json
{ "query": "postgres", "limit": 5 }
```

```json
{ "slug": "example-postgres" }
```

```json
{ "url": "https://www.pryno.ai" }
```

The readiness check fetches public files only (`llms.txt`, `llms-full.txt`, `robots.txt`, OpenAPI, API docs, MCP discovery, and schema.org JSON-LD). It returns pass or fail with a short summary of our own. It does not return excerpts from the target site. Private, loopback, and link-local addresses are rejected.

Resources:

- `pryno://directory/categories`
- `pryno://server/{slug}`

Prompt: `find_mcp_server`.

## Connect

The remote URL for every client is `https://mcp.pryno.ai/mcp`. No API key is required.

### Claude

Add a custom connector with the remote MCP URL `https://mcp.pryno.ai/mcp`.

### ChatGPT

In developer mode, add a remote MCP server with the URL `https://mcp.pryno.ai/mcp`.

### Cursor

```json
{
  "mcpServers": {
    "pryno": {
      "url": "https://mcp.pryno.ai/mcp"
    }
  }
}
```

### VS Code

```json
{
  "servers": {
    "pryno": {
      "type": "http",
      "url": "https://mcp.pryno.ai/mcp"
    }
  }
}
```

### Local stdio

The database URL belongs in the environment that launches the process. Do not commit it.

```json
{
  "mcpServers": {
    "pryno": {
      "command": "npx",
      "args": ["-y", "pryno-mcp"],
      "env": {
        "DATABASE_URL": "postgres://directory_reader@localhost:5432/pryno"
      }
    }
  }
}
```

You can also point the [MCP Inspector](https://modelcontextprotocol.io/docs/tools/inspector) at a local HTTP server:

```bash
npm run build
npm start
npx @modelcontextprotocol/inspector
```

Connect the inspector to `http://localhost:3000/mcp`.

## Self-hosting

The server listens on `PORT` and exposes:

- `GET /` a short description
- `GET /health` database ping and schema status
- `POST /mcp` the Streamable HTTP endpoint
- `GET /mcp/server-card` the MCP server card
- `GET /.well-known/ai-catalog.json` a catalog entry that points at the card

```bash
docker build -t pryno-mcp .
docker run --rm -p 3000:3000 --env-file .env pryno-mcp
```

On Railway, use the included `Dockerfile` and `railway.json`. Keep the project private. Set variables in Railway, not in git.

This process expects the crawler's schema and a `directory_reader` role with `SELECT` only. It does not run migrations. Until those tables exist, `/health` still succeeds when `SELECT 1` works, and the directory tools explain that the schema is not ready. `run_readiness_check` does not need the database.

`readiness_score` and `last_checked_at` are columns the crawler stores on `servers`. If they are missing, the schema is reported as incomplete.

## Environment

Copy `.env.example` to `.env`. Never commit `.env`.

| Variable | Required | Purpose |
| --- | --- | --- |
| `DATABASE_URL` | For directory tools | Postgres URL for the `directory_reader` role. |
| `PORT` | No | HTTP port. Defaults to 3000. Railway sets this. |
| `PUBLIC_URL` | No | Public origin used in the server card. Defaults to `https://mcp.pryno.ai`. |
| `RATE_LIMIT_PER_MIN` | No | Per-IP limit for `/mcp`. Defaults to 60. Readiness checks use a tighter limit. |
| `READINESS_CHECK_URL` | No | When set, `run_readiness_check` POSTs `{ "url" }` to this https endpoint instead of scoring locally. |

## Development

```bash
npm ci
npm test
npm run typecheck
npm run lint
```

Tests cover tool schemas, the search query builder, the SSRF guard, and readiness scoring. They use a mocked database and do not call the network.