Skip to content

Troubleshooting & FAQs ​

What the failures mean, and what to do about each. If you are still stuck, asking your assistant which workspace it is connected to settles most of these in one question.

"Request blocked" before you can sign in ​

A CloudFront error page, rather than a Zefi one, when the browser opens for sign-in:

text
403 ERROR
The request could not be satisfied.
Request blocked.

Your browser is carrying more cookies for zefi.ai than the request can hold — most often because you have signed into more than one Zefi environment, and each leaves its own set behind for thirty days.

Clear your zefi.ai cookies and start the connection again. In Chrome: the icon to the left of the address bar → Cookies and site data → Manage on-device site data → remove the zefi.ai entries. Signing out of Zefi everywhere and back in has the same effect.

This is ours to fix rather than yours to work around, and a fix is in progress.

"Requires re-authorization" or "token expired" ​

The connection is no longer valid. Reconnect through your client — /mcp in Claude Code, the connector entry in Claude, the MCP panel elsewhere.

Usually it means your access to the workspace changed, or the connection has been idle long enough to lapse. Reconnecting is the whole fix.

"Forbidden: no membership for … in …" ​

The connection is fine; the account no longer belongs to that workspace.

Ask your assistant which workspace it is connected to, and compare it with the one you expect. If you have been moved between workspaces, reconnect and choose the current one — a connection cannot be repointed, because the workspace is fixed when you approve it.

"That workspace requires signing in through its SSO provider" ​

The workspace enforces SSO, and this session signed in another way — with Google or a password, for example.

Sign in to Zefi through that workspace's SSO first, then connect again. Signing in some other way will not do, however valid your membership: enforcement is the point of the setting.

A workspace is missing from the list ​

Two possibilities, in the order worth checking:

  1. You are not a member. The list is the same one the platform shows you. Whoever administers the workspace can add you.
  2. It enforces SSO and you did not use it. Sign out, sign back in through that workspace's SSO, and start again.

The client connects but finds no tools ​

Almost always the client rather than the connection.

  • Confirm the URL ends in /mcp — https://mcp.zefi.ai alone is not the server.
  • In VS Code, the chat must be in Agent mode; tools are not offered in Ask mode.
  • In ChatGPT, add the connection to the conversation from the tools menu.
  • Some clients cache the tool list. Restart, or toggle the server off and on.

Then ask your assistant to name a Zefi tool and use it directly. If that works, the connection is healthy and the question is how your client surfaces tools.

Sign-in opens, but never completes ​

Usually the browser rather than us.

  • Finish in the window the client opened; a copied link in another profile signs in as a different account.
  • The sign-in page is expected — Zefi's own login, not a Zefi-branded page on a domain you do not recognise. Only mcp.zefi.ai and login.zefi.ai are ours.
  • Through mcp-remote, clear ~/.mcp-auth and retry.
  • A sign-in left more than ten minutes ends and has to be started again.

Checking the server itself ​

The server publishes its own metadata, so you can confirm it is reachable and that the problem is nearer your client:

bash
curl https://mcp.zefi.ai/.well-known/oauth-protected-resource/mcp
json
{
  "resource": "https://mcp.zefi.ai/mcp",
  "authorization_servers": ["https://mcp.zefi.ai"],
  "scopes_supported": ["default"],
  "bearer_methods_supported": ["header"],
  "resource_name": "Zefi MCP"
}

A 401 from https://mcp.zefi.ai/mcp is also correct rather than a fault — it is how the server tells a client where to sign in.

Disconnecting ​

Remove the server in your client, which both stops it being offered and ends the connection. In Claude Code, claude mcp remove zefi. Ending a connection cannot be undone — reconnecting creates a new one, and asks you to approve again.

Common questions ​

Can I connect more than one workspace?

Yes — add the server again under a different name and approve the other workspace. A connection cannot be repointed, so a second one is how you reach a second workspace.

bash
claude mcp add --transport http zefi-acme https://mcp.zefi.ai/mcp
claude mcp add --transport http zefi-globex https://mcp.zefi.ai/mcp

Through the bridge, give each its own credential directory, or the second sign-in overwrites the first:

json
{
  "env": { "MCP_REMOTE_CONFIG_DIR": "~/.mcp-auth/zefi-acme" }
}
Can I use a Zefi API key instead of signing in?

No. An API key presented as a bearer token is refused — the server accepts only credentials it issued itself.

A key acts as itself with fixed scopes. An MCP connection acts as you, with your permissions, checked on every call. Keys stay right for the REST API.

Is there an SSE endpoint?

No. Streamable HTTP at https://mcp.zefi.ai/mcp, and nothing else. A client that speaks only stdio should use the bridge.

Do I need to install anything?

Not for a client with native remote support — the URL is the whole setup. The bridge needs Node.js 18 or later, which is its requirement rather than ours.

What can the assistant reach?

What you can reach, in the workspace you approved, and nothing outside it. Permissions are read on every call rather than fixed when you connect.

What happens if I leave the workspace?

It stops working at the next call. Nothing to revoke and no delay to wait out.

Zefi API v1