Beta. The assistant API and its reads over MCP are in beta. They are open to every
organization that pays out through Grade, and parts are not available yet, as each page says.
https://mcp.usegrade.com/mcp. People in your organization also find it in Grade, under
Account settings → AI agents, and on the Developer page, beside the API base URL.
During the beta, every other organization is refused with 403, and where the server is off it
answers 404.
Your own model
Mint an agent token on your backend, as for a question (Quickstart), and send it as the bearer:
With Claude’s MCP connector,
pass the server and the token with each request:
Claude, for a person
A person connects Claude with their own Grade sign-in, and no SDK key is involved. Claude then reads as that person, within what their role reaches, for up to 30 days, and they can disconnect it in Account settings → AI agents. The steps, with screenshots, are in the Help Center. ChatGPT comes later.What it lists
The assistant’s own reads, within the token’s scopes: see Reads. Each answers at most 20 rows, with the full count. Actions are never listed: a change goes through the assistant, as a proposal a person confirms.The protocol
Streamable HTTP without sessions, answered in JSON: one JSON-RPC message perPOST, in the protocol
revisions 2025-11-25, 2025-06-18 and 2025-03-26. A read that fails comes back as a tool result
with isError: true and a sentence your model can act on.