Skip to content
Agent Fleet

Connect a remote MCP client by signing in

claude.ai, ChatGPT, Claude Code and Codex can connect to your workspace without a token. Sign in to the Console from the client and allow it on the consent page; that is all.

How it works

Section titled “How it works”
What you do What the client does
Add your workspace’s address to the client Opens the Console’s sign-in and the consent page
Choose a tenant and a scope on the consent page, and allow Receives a short-lived access token and calls the MCP tools, renewing it on its own
Revoke a connection you no longer need Is refused from its next call

Connect

Section titled “Connect”

  1. In the client, add a custom connector or remote MCP server, with one of these addresses. Leave any client ID and secret fields empty

    Address What it reaches
    https://<your Console address>/mcp The session tools and memory
    https://<your Console address>/mcp/memory Agent Fleet memory only

  2. The client opens the Console. Sign in if you are not signed in, and the consent page opens

  3. Check the consent page (table below) and press “Allow”. To decline, press “Cancel”: the app is told you refused

To connect Claude Code, register the server for your user, then choose /mcp → agent-fleet → Authenticate inside Claude Code.

Terminal window
claude mcp add --transport http --scope user agent-fleet https://<your Console address>/mcp

Check the consent page

Section titled “Check the consent page”

The consent page opens under the heading “An app wants to connect to Agent Fleet”, or “An app wants to connect to your Agent Fleet memory” for /mcp/memory.

Field What to look at
After connecting, you are sent back to The host in large type (claude.ai, chatgpt.com, or “this computer itself” for a command-line client). This is what tells you whether you started the connection
Registered by A registration checked against a document the app’s publisher published (CIMD), or dynamic client registration (DCR), which anyone can do
App name Marked “declared by the app itself, not verified”. The app chose it freely, so do not rely on it
Tenant to connect One of the tenants you belong to. The connection is fixed to it
What to allow read or write

The scopes are:

Scope What it allows
read Read session status and output. List, search and read memory
write Read, plus act: send to sessions and similar. Save and forget memory
  • read starts selected. write is offered only when the app asked for write. You can always choose less than the app asked for
  • Memory is reachable only once you have allowed outside use of memory and your tenant administrator has not blocked it
  • admin:dangerous is never granted this way. Use a token if you need it

Once you allow it, the client receives an access token that expires after an hour and renews it on its own to stay connected. You are not asked again.

How long a connection lasts

Section titled “How long a connection lasts”
It ends when Then
It has not been used for 30 days The app asks you to sign in and allow it again
90 days have passed since you allowed it Same
You leave the tenant It is refused from the next call
A tenant administrator turns MCP OAuth off Every connection in that tenant is revoked and no new one can be made. Tokens keep working

What a connection does is recorded in the audit log as an action over MCP.

Check and revoke connected apps

Section titled “Check and revoke connected apps”

In ⚙Settings → “MCP tokens & apps”, “Connected apps” appears below the token list. The section is shown only when the deployment offers MCP OAuth.

The "Connected apps" section: a read connection from claude.ai to /mcp/memory and a write connection to /mcp from this computer itself (127.0.0.1), with the second one's "Recent calls" open on tools/call, tools/list and other requests
Column Contents
App The app’s name, marked “Self-declared name (unverified)”
Redirect host Where you were sent back to when you allowed it
Tenant The tenant it is connected to
Endpoint /mcp (full) or /mcp/memory (memory only)
Scope read or write
Connected and Last used The day you allowed it, and the day it was last used

Press “Revoke” and confirm, and the connection stops working at once, refresh token included. To use it again, connect again from the app.

When it goes wrong

Section titled “When it goes wrong”

Connected but no tools appear

Section titled “Connected but no tools appear”

Open “Recent calls” on a connected app’s or a token’s row to see the last 50 requests that came in on that connection.

Column Contents
Time and Status When it arrived, and the HTTP method and status
JSON-RPC The method (per element in a batch). The number of tools returned for tools/list, or the error code
Protocol The protocol era and version
Client The client’s name and version
Headers MCP-Protocol-Version, whether Mcp-Method / Mcp-Name were sent, and Accept
Latency The time to answer
  • Tokens, authorization headers, and tool arguments and results are not recorded
  • The list lives only in the Control Plane’s memory and is lost on restart
  • With several Control Planes, each keeps its own list, so you see only what the one that answered received. Press “Refresh” a few times, or ask an operator to check the Control Plane’s log: it records every request, including those refused before sign-in

An error page appears instead of the sign-in

Section titled “An error page appears instead of the sign-in”

Only addresses the deployment approved may receive the sign-in result. By default those are claude.ai, claude.com, chatgpt.com, and command-line clients on your own computer (127.0.0.1, localhost). A client asking to be sent back anywhere else is refused with an error page, and no code is sent. To use another client, talk to your administrator.

A connection that used to work asks you to sign in

Section titled “A connection that used to work asks you to sign in”

It was not used for 30 days, or 90 days have passed since you allowed it. Follow the app’s prompt, sign in and allow it again.

Section titled “Related”