Connect an AI assistant to Nod

Authorization is discovered from it: the endpoint answers an unauthenticated call with a challenge naming these two documents, and every conforming client follows them without being told.

https://nod.byrewired.com/.well-known/oauth-protected-resource/api/mcp
https://nod.byrewired.com/.well-known/oauth-authorization-server

Protocol revisions 2025-03-26 through 2026-07-28 are all served from that one URL. A request that omits the version header is read as 2025-03-26, which is what the specification says to do.

claude mcp add --transport http --scope user nod https://nod.byrewired.com/api/mcp

https://nod.byrewired.com/api/mcp

https://nod.byrewired.com/api/mcp

{
  "mcpServers": {
    "nod": {
      "url": "https://nod.byrewired.com/api/mcp",
      "headers": {
        "Authorization": "Bearer nod_sk_your_key_here"
      }
    }
  }
}

{
  "servers": {
    "nod": {
      "type": "http",
      "url": "https://nod.byrewired.com/api/mcp",
      "headers": {
        "Authorization": "Bearer nod_sk_your_key_here"
      }
    }
  }
}

{
  "context_servers": {
    "nod": {
      "source": "custom",
      "url": "https://nod.byrewired.com/api/mcp",
      "headers": {
        "Authorization": "Bearer nod_sk_your_key_here"
      }
    }
  }
}

Make a key in Settings, then AI connections, and paste it in place of the placeholder. A key acts as whoever made it and works until it is revoked, so keep it out of anything you share.

An assistant that needs more asks again: a tool it cannot reach answers with the exact permission it wants, and the app can send you back to the approval screen for that one thing.

You can always grant less than an app asks for. A connection approved with read only stays read only, whatever the app requested, and nothing about it changes until you approve it again.

The names beside a tool are the arguments it insists on. Everything else it takes is optional and has a sensible default.

• Approve, reject, sign or accept. There is no tool for it at any permission level, and there will not be: an approval is a recorded decision by a named person, and in this market it is increasingly the document that triggers payment. An assistant can read the whole thread and draft the reply. A person presses the button.
• Return your work. No tool, resource or prompt hands back a frame, a render or a file. Delivery stays behind the watermarking and expiring links the product already uses.
• See past the person who approved it. Every query runs as that account under the same rules the browser enforces, so an assistant connected by a client sees a client's view, not the agency's.
• Send anything to anyone. It cannot issue a review link, invite a person, email a client or change who can see a project. Reaching outside the workspace stays a decision somebody makes in the product.

120 calls a minute per connection, counted centrally rather than per server, so the limit is the limit. Past it, calls come back as 429 and the assistant waits.

Every call is recorded with the tool it used, the account it acted as, whether it succeeded and how long it took. Workspace admins can read their own workspace's log; support staff cannot, because tool arguments carry project and asset names.

Disconnecting is immediate. Revoking a connection invalidates every token issued under it in the same statement, so there is no hour-long window where an assistant keeps working.

Access tokens are short lived and refreshed in the background. A refresh token is single use: presenting one twice means two parties hold it, so the whole connection is revoked rather than the second attempt refused.

Anything else, open Settings, then AI connections. Every call the workspace has made is there with its outcome, which is usually enough to see what the app is actually asking for.