Polygraph MCP server

Polygraph is a Model Context Protocol (MCP) server. An AI agent connected to it can ask a person or a business for verified answers on its user’s behalf, and can show its user the requests they receive so they can agree, suggest a change or decline. This page is the documentation for developers and agents connecting to it.

Connect

Server address
https://mcp.polygraphagent.com/mcp
Transport
Streamable HTTP
Sign-in
OAuth 2.0, authorization code with PKCE (S256). Every request needs a token; there is no API key.
Scopes
openid, profile, email, offline_access
Client configuration
/.well-known/mcp.json

Add it to an AI agent

  1. ChatGPT or CodexOpen Plugins, choose Add, then Create custom MCP server. Paste the server address as the Server URL and keep Authentication on OAuth.
  2. ClaudeOpen Settings, then Connectors. Add a custom connector and paste the server address.
  3. Any other MCP clientPoint it at the server address. An unauthenticated request is answered 401 with a WWW-Authenticate header naming the resource metadata, from which the client finds the authorization server.

Open setupInstructions for agents

What a connected agent can do

Ask
Pass on what the user wants to know and who to ask. Polygraph drafts a title and a few short numbered questions, each with its answer form, any definitions and its evidence criteria. Nothing is sent until the user approves the exact questions.
Read and decide
Show the user a request they received, in the same words the requester approved, and record their own decision: agree to the whole request, suggest a change, or decline.
Answer
After the user agrees, look through their records privately, show them every answer first, and send the answers they approve, all together. Not every AI agent can send answers yet; when a request is opened, the agent is told whether it can.
The user decides
Approval, agreement and the go-ahead to send answers are the user’s own. An agent records them and never makes them.

This site, for agents

Every page is available as Markdown: request it with Accept: text/markdown. llms.txt says when to use Polygraph and how. The agent view gives the same facts as JSON.