In plain words: an MCP server can tell clients "sign in with this login service" while not needing a sign-in at all. Some clients ignore the sign. Others obey it, can't complete the login, and quietly show the user no tools.
I test how coding agents use a vendor's docs channels, one channel at a time. One of the channels was Kestra's hosted docs MCP at https://api.kestra.io/v1/mcp. Claude Code connected and called it. In Codex the server's tools never showed up. There was no error in the log either, because Codex hadn't loaded them at all.
Here's what the server does, all checkable with curl:
- It answers
initializeandtools/listwithout any token: 12 tools, open to anyone. - At the same time it publishes OAuth protected-resource metadata that names
https://accounts.google.comas its authorization server. - Google's metadata has no
registration_endpoint. A client can't register itself there, so a client without a pre-made client ID can't log in.
Claude Code connects without a token, gets 200, and lists the tools. Codex reads the metadata as "login required". codex mcp list shows the server as Not logged in, and the session has no tools from it. In my first probe, codex mcp add failed with "Dynamic client registration not supported". I re-ran both clients today with the same result.
What the docs don't say: a server that is fully open still loses every client that takes its OAuth metadata at its word. The metadata is a promise. If it points to a login no client can finish, the server is only as open as the least strict client.
The opposite mistake is in my older story My MCP server hid its own tool list behind a login. That one is about tools that needed a login nobody could see. This is the same check from the other side, and it's now part of how I test MCP servers before comparing them.
For the experiment, Codex reached the same server through a small stdio bridge with no auth. A Codex user following Kestra's setup page wouldn't have that bridge.
Check if this applies to you
Applies if: you run, or connect agents to, a remote MCP server over HTTP (Streamable HTTP). It doesn't apply to stdio servers.
Read-only steps, with URL as the MCP endpoint (for example https://api.example.com/mcp) and ORIGIN as its scheme and host:
curl -s "$ORIGIN/.well-known/oauth-protected-resource<path of URL>"andcurl -s "$ORIGIN/.well-known/oauth-protected-resource". If both return 404 or no JSON, the server advertises no OAuth: it doesn't have this problem, so stop here.- For each entry in
authorization_servers, fetch<server>/.well-known/oauth-authorization-server(or/.well-known/openid-configurationif that 404s). Note whether it has aregistration_endpointor"client_id_metadata_document_supported": true. - Without any
Authorizationheader, POST{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"check","version":"0"}}}, then{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}, toURLwithContent-Type: application/jsonandAccept: application/json, text/event-stream. Note the status code and the number of tools.
Has the problem: step 1 returns metadata, and step 2 shows no way to register (no registration_endpoint, no client ID metadata support). If step 3 also lists tools anonymously, it's the mixed case from this story: open tools that strict clients like Codex hide. If step 3 returns 401, no client without a pre-registered client ID can connect at all. Doesn't have it: step 1 finds no metadata, or step 2 shows a registration path.
Reproduce this
Prerequisites: Claude Code and Codex CLI installed and signed in; a remote MCP server URL to test. Your agent will need from you: the MCP server URL. Steps:
- Run the three read-only checks above and write down what they return.
- Claude Code, for one run without touching your config:
claude -p "List the exact tool names you have from the MCP server named probe, or say none. Do not call any tool." --strict-mcp-config --mcp-config '{"mcpServers":{"probe":{"type":"http","url":"<URL>"}}}' - Codex, also only for this run:
codex -c 'mcp_servers.probe.url="<URL>"' mcp list(read the Auth column), thencodex exec --skip-git-repo-check -s read-only -c 'mcp_servers.probe.url="<URL>"' "List the exact tool names you have from the MCP server named probe, or say none. Do not use the shell or the web."
Verify: if the server is in the mixed case, Claude Code lists the tools while Codex shows Not logged in and answers "none". If the two clients list the same tools, the server doesn't have this problem.
What I'd change as the server owner (my reading, not tested on Kestra's server): if the tools are meant to be public, don't publish protected-resource metadata for the MCP path. If they need a login, point the metadata at an authorization server that clients can register with, through dynamic client registration or client ID metadata documents.