Supported MCP revisions
The gateway speaks these MCP revisions, with AI clients and with upstream servers:2026-07-28 · 2025-11-25 · 2025-06-18 · 2025-03-26 · 2024-11-05
The OAuth behavior on this page is the same whichever revision a client or server speaks.
What is supported, by leg
How an AI client identifies itself to the gateway
MCP Manager tells the two methods apart by theclient_id the client sends. If it is an https URL, MCP Manager treats it as a client ID metadata document, and refuses it if it isn’t a usable document URL. Any other client_id is looked up as a registered client, for example one that registered dynamically.
Client ID metadata documents
The client publishes a JSON document at a URL it controls and uses that URL as itsclient_id. MCP Manager fetches the document and checks it before the user sees a consent screen. The document must meet these requirements:
- Its URL, which is the
client_id, ishttps, has a path longer than/, has no embedded credentials or control characters, and is at most 2,048 characters long. - It is served over
https, straight from the URL. MCP Manager doesn’t follow redirects, so a document behind anhttptohttpsredirect or a vanity-domain redirect is refused. - It is served as
application/json, is no larger than 64 KiB, and answers within 5 seconds. - It is served from a publicly reachable address. MCP Manager refuses a document on a private network address.
- Its
client_idfield equals the URL it was fetched from. - It has a
client_nameof at most 255 characters and aredirect_urislist of at most 20 entries, each at most 2,048 characters long. - Each redirect URI passes the same checks MCP Manager applies when a client registers dynamically. For example, no
javascript:scheme, no embedded credentials, andhttponly on a loopback address. - It declares a
token_endpoint_auth_methodofnoneorprivate_key_jwt. A document that leaves it out is refused. - If it has a
logo_urior aclient_uri, each is anhttpsURL of at most 2,048 characters.
redirect_uri the client sends at sign-in must match one the document lists, character for character. The one exception is the port on an http loopback address such as http://127.0.0.1 or http://localhost: a document may list one without a port, and the client may then send the port it opened.
When MCP Manager refuses a document, the error response gives the reason. If the request came from a signed-in workspace member, an admin may also see an AI client sign-in alert.
Signed client assertions
A client document can declareprivate_key_jwt and publish its keys with jwks_uri or an inline jwks. MCP Manager registers the client for private_key_jwt when it can use those keys. For such a client, MCP Manager verifies the signed assertion with the keys the client publishes, at the token, revocation, and introspection endpoints. MCP Manager supports these algorithms: RS256, RS384, RS512, PS256, PS384, PS512, ES256, ES384, ES512, and EdDSA.
The assertion’s audience must be the URL of MCP Manager’s token, revocation, or introspection endpoint, or its issuer URL. It must expire within one hour (plus about a minute of clock skew) and carry a unique jti. MCP Manager accepts each assertion only once. If MCP Manager then refuses the request the assertion signed, the assertion can be used again. A client that registered dynamically can use private_key_jwt too.
How MCP Manager identifies itself to an upstream server
MCP Manager first discovers the server’s authorization server. It reads the server’s protected resource metadata, then the authorization server’s own metadata, from well-known paths. If that doesn’t work, MCP Manager looks for authorization server metadata at the MCP server’s own address. If all of that fails, it follows theresource_metadata URL in the server’s 401. It then picks how to identify itself:
- A client that already exists for the server is normally reused. That is a client you pre-registered, or one MCP Manager already holds for the server.
- Otherwise, a client ID metadata document, when the authorization server advertises support for them. MCP Manager does not register. It presents the address of its own document as its
client_id. - Otherwise, dynamic client registration, at the registration endpoint the authorization server advertises. When none is advertised, MCP Manager tries
/registeron the MCP server’s own address: its scheme, host, and port, without the path.
MCP Manager's client ID metadata document
client_id is that same URL. It lists the callback URL as its only redirect URI. It declares token_endpoint_auth_method as none, because MCP Manager is a public client that uses PKCE, and it publishes no keys.
With a client ID metadata document, the upstream authorization server fetches the document from MCP Manager over the internet. If the authorization
server’s network blocks outbound requests or uses an allowlist, allow that URL, or the authorization server cannot identify MCP Manager.
Resource indicators sent upstream
When MCP Manager authorizes against an upstream server, it reads the server’s protected resource metadata again. This time it tries theresource_metadata URL from the server’s 401 first, and the well-known path after that. It then decides which resource indicator to send:
- If the server’s protected resource metadata names a usable
resourceon the server’s own origin, MCP Manager sends it. - If the metadata names no
resource, MCP Manager sends the server’s own address, without any query, fragment, or credentials. - If there is no metadata, the named value is unusable or on another origin, or the server’s address isn’t
https, MCP Manager sends nothing. - If the lookup itself fails (for example, it times out) and the server was authorized before, MCP Manager resends the resource from the server’s most recently updated grant, as long as it is still usable. When MCP Manager knows the server’s address, that resource must also be on the server’s origin. If nothing usable is stored, it sends nothing.
access_denied, login_required, consent_required, interaction_required, invalid_client, invalid_grant, server_error, and temporarily_unavailable, and also a bare HTTP 400. It doesn’t retry on a 5xx.
What is not supported
- OpenID Connect discovery. The gateway does not serve
/.well-known/openid-configuration, and it does not use it to find an upstream server’s endpoints. It uses RFC 8414 and RFC 9728 metadata only. - Grant types other than authorization code and refresh token. The gateway’s token endpoint does not issue tokens for the
client_credentialsgrant or the device authorization grant. - Signed assertions from MCP Manager to an upstream server. MCP Manager identifies itself upstream with a client secret or as a public client with PKCE. It does not use
private_key_jwtand publishes no keys. - Upstream servers that accept only clients they have approved. MCP Manager cannot connect to such a server until its vendor approves MCP Manager. Figma’s remote server is an example. See Connect Figma.
Further reading
Authentication & Identity
The two separate authentications, and how MCP Manager stores and refreshes the credentials it brokers.
Choosing a Remote Server's Authentication Method
When to use standard OAuth, client pre-registration, or a token in custom headers.
Alerts
The alerts an admin sees when an AI client cannot be verified.
External sources
MCP authorization spec
The OAuth-based authorization model defined by the Model Context Protocol.
RFC 7523
JWT profile for client authentication, the basis of
private_key_jwt.
