@polygraphso/mcp

MCP server for polygraph.so — independent, lab-evaluated trust grades for MCP servers, exposed as native tools to any MCP client (Claude Desktop, Cursor, and others).

A polygraph is the result polygraph.so issues for an MCP server: an adoption tier (Top 10 / 25 / 50 / 100), and — once behavioral evaluation has run — an A–F grade plus an evidence URL. This package lets an agent check the polygraph for a server before recommending or installing it.

Tools

• check_server — look up the polygraph for a specific MCP server (server_ref like npm/@modelcontextprotocol/server-filesystem).
• list_servers — enumerate every server polygraph tracks, tier-sorted.

notify_about (request a polygraph for an untracked server) lands in v0.2 — the underlying endpoint isn't published yet.

Install in Claude Desktop

Edit claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "polygraph": {
      "command": "npx",
      "args": ["-y", "@polygraphso/mcp"]
    }
  }
}

Restart Claude Desktop. The two polygraph tools appear in the tool picker.

Try it:

Use polygraph to check whether npm/@modelcontextprotocol/server-filesystem is safe to install.

Install in Cursor

Cursor reads MCP servers from ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project):

{
  "mcpServers": {
    "polygraph": {
      "command": "npx",
      "args": ["-y", "@polygraphso/mcp"]
    }
  }
}

Reload the Cursor window after editing.

Install in any other MCP client

The server speaks stdio. Anything that can launch a stdio MCP server with npx -y @polygraphso/mcp will work. The binary name is polygraphso-mcp if you want to invoke it directly.

Local install

npm i -g @polygraphso/mcp
polygraphso-mcp   # talks JSON-RPC on stdin/stdout

What it returns

check_server({ server_ref: "npm/lodash" }):

{
  "status": "tracked",
  "adoption_tier": "top10",
  "polygraph": null,
  "notify_url": "https://polygraph.so/notify?for=npm/lodash"
}

• status is "tracked" if polygraph is evaluating the server, "not_available" otherwise.
• adoption_tier is top10 / top25 / top50 / top100, or null if the server is tracked but unranked.
• polygraph is null until the behavioral evaluation lands (v0 has the adoption side; the litmus harness is shipping next).
• notify_url is where a user can subscribe to be notified when the polygraph is published.

list_servers():

{
  "servers": [
    {
      "server_ref": "npm/@modelcontextprotocol/server-filesystem",
      "adoption_tier": "top10",
      "polygraph": null
    }
  ],
  "total": 75
}

Honest coverage

A server with no polygraph yet is neither safe nor unsafe — it's unevaluated. The tool surfaces this directly; agents using this MCP should pass that distinction through to the user.

Configuration

| Env var | Default | Purpose | |---|---|---| | POLYGRAPH_API_URL | https://polygraph.so | Override the API base, e.g. for local development (http://localhost:3000). |

Links

• Site: https://polygraph.so
• CLI: https://www.npmjs.com/package/polygraphso (npx polygraphso check <ref>)
• Source: https://github.com/polygraphso/core
• Issues: https://github.com/polygraphso/core/issues

License

Apache-2.0