From 229e13678320cc18fc496ed4bfe7f6af601f6477 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 11:26:54 -0700 Subject: [PATCH 1/6] fix(mcp): honor admin-entered OAuth URLs on authorize after issuer yield Co-authored-by: Cursor --- .../mcp_server/discoverable_endpoints.py | 55 +++++++++----- .../mcp_server/mcp_server_manager.py | 6 +- .../types/mcp_server/mcp_server_manager.py | 12 ++++ .../mcp_server/test_discoverable_endpoints.py | 72 +++++++++++++++++++ 4 files changed, 126 insertions(+), 19 deletions(-) diff --git a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py index aef4f5dc721..6ede1c553e8 100644 --- a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py +++ b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py @@ -663,6 +663,23 @@ def _endpoint_not_configured_detail( ) +async def _server_with_oauth_endpoints(mcp_server: MCPServer) -> MCPServer: + """Join deferred OAuth discovery only when this server still has no authorize URL. + + Admin-entered endpoints live on ``configured_*`` after an anchored issuer empties the + resolved fields. Those already let authorize/token run, so discovery is not awaited + and cannot 503 over a leftover pin. A server with nothing configured still joins the + deferred task; no slot is a no-op and the caller 400s. + """ + if mcp_server.effective_authorization_url is not None: + return mcp_server + from litellm.proxy._experimental.mcp_server.mcp_server_manager import ( # noqa: PLC0415 # circular import with mcp_server_manager at module load + global_mcp_server_manager, + ) + + return await global_mcp_server_manager.ensure_oauth_metadata_discovered(mcp_server) + + def _raise_unless_oauth2_discovery_server( mcp_server: MCPServer | None, mcp_server_name: str | None, @@ -697,7 +714,7 @@ def _dcr_bridge_relays_client_registration(mcp_server: MCPServer) -> bool: returns directly to the client's redirect URI without transiting the gateway. Gateway-side redirect trust and the ``/callback`` state relay therefore only apply to the short-circuit arm, where the upstream only knows the gateway's own callback.""" - return mcp_server.is_dcr_bridge and bool(mcp_server.registration_url) and not mcp_server.client_id + return mcp_server.is_dcr_bridge and bool(mcp_server.effective_registration_url) and not mcp_server.client_id def _require_s256_pkce( @@ -745,7 +762,7 @@ def _redirect_to_upstream_authorize( **({"scope": scope_value} if scope_value else {}), **({"resource": upstream_resource} if upstream_resource else {}), } - parsed_auth_url: Final = urlparse(mcp_server.authorization_url or "") + parsed_auth_url: Final = urlparse(mcp_server.effective_authorization_url or "") merged_params: Final = {**dict(parse_qsl(parsed_auth_url.query)), **passthrough_params} return RedirectResponse(urlunparse(parsed_auth_url._replace(query=urlencode(merged_params)))) @@ -812,11 +829,12 @@ async def authorize_with_server( ephemeral_dcr_client: "EphemeralDcrClient | None" = None, ): _raise_if_not_oauth2(mcp_server) - if mcp_server.authorization_url is None: + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server) + if resolved_server.effective_authorization_url is None: raise HTTPException( status_code=400, detail=_endpoint_not_configured_detail( - mcp_server, + resolved_server, "authorization url", "set Authorization URL and Token URL manually", "set Issuer to discover them from the identity provider (RFC 8414)", @@ -913,7 +931,7 @@ async def authorize_with_server( if upstream_resource: params["resource"] = upstream_resource - parsed_auth_url: Final = urlparse(mcp_server.authorization_url) + parsed_auth_url: Final = urlparse(resolved_server.effective_authorization_url) existing_params: Final = dict(parse_qsl(parsed_auth_url.query)) existing_params.update(params) final_url: Final = urlunparse(parsed_auth_url._replace(query=urlencode(existing_params))) @@ -946,11 +964,13 @@ async def exchange_token_with_server( if grant_type not in ("authorization_code", "refresh_token"): raise HTTPException(status_code=400, detail="Unsupported grant_type") - if mcp_server.token_url is None: + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server) + token_url: Final = resolved_server.effective_token_url + if token_url is None: raise HTTPException( status_code=400, detail=_endpoint_not_configured_detail( - mcp_server, + resolved_server, "token url", "set Token URL manually", "set Issuer to discover it from the identity provider (RFC 8414)", @@ -1067,7 +1087,7 @@ async def exchange_token_with_server( async_client: Final = get_async_httpx_client(llm_provider=httpxSpecialProvider.Oauth2Check) try: response: Final = await async_client.post( - mcp_server.token_url, + token_url, headers={"Accept": "application/json", **token_request.headers}, data=token_data, ) @@ -1551,7 +1571,8 @@ async def mint_ephemeral_dcr_client(request: Request, mcp_server: MCPServer) -> bounded by the server count even when the request origin varies) so parallel authorize requests cannot each register an upstream client; the cache stamps nothing onto the server record and correctness never depends on it because the sealed state carries the client through the flow.""" - if mcp_server.registration_url is None: + registration_url: Final = mcp_server.effective_registration_url + if registration_url is None: return None request_base_url: Final = get_request_base_url(request) cache_key: Final = f"mcp_ephemeral_dcr_client:{mcp_server.server_id}:{request_base_url}" @@ -1571,7 +1592,7 @@ async def mint_ephemeral_dcr_client(request: Request, mcp_server: MCPServer) -> "token_endpoint_auth_method": "none", } response: Final = await _post_dcr_registration( - registration_url=mcp_server.registration_url, + registration_url=registration_url, register_data=register_data, server_id=mcp_server.server_id, ) @@ -1617,7 +1638,7 @@ async def resolve_ephemeral_dcr_client( usable to generate orphan IdP clients).""" if not (mcp_server.is_true_passthrough or (mcp_server.is_oauth_delegate and not mcp_server.is_dcr_bridge)): return None - if mcp_server.authorization_url is None: + if mcp_server.effective_authorization_url is None: raise HTTPException( status_code=400, detail="MCP server authorization url is not set", @@ -1661,21 +1682,23 @@ async def register_client_with_server( ): return dummy_return - if mcp_server.authorization_url is None: + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server) + if resolved_server.effective_authorization_url is None: raise HTTPException( status_code=400, detail=_endpoint_not_configured_detail( - mcp_server, + resolved_server, "authorization url", "set Authorization URL and Token URL manually", "set Issuer to discover them from the identity provider (RFC 8414)", ), ) - if mcp_server.registration_url is None: + registration_url: Final = resolved_server.effective_registration_url + if registration_url is None: return dummy_return - bridge_relay: Final = _dcr_bridge_relays_client_registration(mcp_server) + bridge_relay: Final = _dcr_bridge_relays_client_registration(resolved_server) if bridge_relay and not client_redirect_uris: raise HTTPException( status_code=400, @@ -1690,7 +1713,7 @@ async def register_client_with_server( "token_endpoint_auth_method": token_endpoint_auth_method or ("none" if bridge_relay else ""), } response: Final = await _post_dcr_registration( - registration_url=mcp_server.registration_url, + registration_url=registration_url, register_data=register_data, server_id=mcp_server.server_id, ) diff --git a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py index 7ab26db0f3e..fe0c73a5efc 100644 --- a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py +++ b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py @@ -523,7 +523,7 @@ def _oauth_endpoints_unresolved(server: MCPServer) -> bool: # can come from resource discovery, so a server that resolved its endpoints but no scopes is # still unresolved for its flow. return True - if server.is_dcr_bridge and not server.client_id and server.registration_url is None: + if server.is_dcr_bridge and not server.client_id and server.effective_registration_url is None: # A DCR bridge with no admin-configured client can only register callers through the # upstream's registration endpoint, so a build that resolved the authorize and token # endpoints but not registration_endpoint (partial metadata) is still unresolved for its @@ -535,8 +535,8 @@ def _oauth_endpoints_unresolved(server: MCPServer) -> bool: return _flow_endpoints_missing( server.auth_type, MCPServerManager.effective_oauth2_flow(server), - server.authorization_url, - server.token_url, + server.effective_authorization_url, + server.effective_token_url, server.token_exchange_endpoint, ) diff --git a/litellm/types/mcp_server/mcp_server_manager.py b/litellm/types/mcp_server/mcp_server_manager.py index d09503cdc4d..401793a79e4 100644 --- a/litellm/types/mcp_server/mcp_server_manager.py +++ b/litellm/types/mcp_server/mcp_server_manager.py @@ -183,6 +183,18 @@ class MCPServer(BaseModel): def __str__(self) -> str: return self.__repr__() + @property + def effective_authorization_url(self) -> str | None: + return self.authorization_url or self.configured_authorization_url + + @property + def effective_token_url(self) -> str | None: + return self.token_url or self.configured_token_url + + @property + def effective_registration_url(self) -> str | None: + return self.registration_url or self.configured_registration_url + @property def has_client_credentials(self) -> bool: """True if this server should use the OAuth2 client_credentials (M2M) flow. diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py index bcac27a4a14..790b5c7ad22 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py @@ -8854,6 +8854,78 @@ async def test_authorize_wall_names_the_issuer_for_anchored_servers(): assert "idp.example.com" not in detail_text +@pytest.mark.asyncio +async def test_authorize_uses_admin_entered_github_oauth_urls_after_issuer_yield(): + """GitHub MCP servers store Authorization URL and Token URL on the row. 1.99 can empty + the resolved authorization_url when a leftover issuer is treated as a pin (RFC 8414 + yield). The UI authorize must still redirect to the admin-entered GitHub authorize URL + instead of 400ing that discovery against api.githubcopilot.com failed.""" + from litellm.proxy._experimental.mcp_server.discoverable_endpoints import ( + authorize_with_server, + ) + from litellm.types.mcp import MCPAuth, MCPTransport + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + server = MCPServer( + server_id="ecac50c4-8eca-438a-af80-9bdebadafc69", + name="github_mcp", + alias="github_mcp", + server_name="github_mcp", + url="https://api.githubcopilot.com/mcp/", + transport=MCPTransport.http, + auth_type=MCPAuth.oauth2, + oauth2_flow="authorization_code", + client_id="github-app-client", + authorization_url=None, + token_url=None, + issuer="https://github.com", + issuer_is_anchored=True, + configured_authorization_url="https://github.com/login/oauth/authorize", + configured_token_url="https://github.com/login/oauth/access_token", + ) + mock_request = MagicMock() + mock_request.base_url = "https://litellm.example.com/" + mock_request.headers = {} + + with patch("litellm.proxy._experimental.mcp_server.discoverable_endpoints.encrypt_value_helper") as mock_encrypt: + mock_encrypt.return_value = "mocked_encrypted_state" + response = await authorize_with_server( + request=mock_request, + mcp_server=server, + client_id="github-app-client", + redirect_uri="http://127.0.0.1:60108/callback", + state="state123", + ) + + assert response.status_code == 307 + assert "https://github.com/login/oauth/authorize" in response.headers["location"] + assert "client_id=github-app-client" in response.headers["location"] + + +def test_oauth_endpoints_count_admin_entered_urls_as_resolved(): + """A leftover issuer empties the resolved authorize/token fields but must not keep the + server on the deferred-discovery retry path when the admin already stored those URLs.""" + from litellm.proxy._experimental.mcp_server.mcp_server_manager import ( + _oauth_endpoints_unresolved, + ) + from litellm.types.mcp import MCPAuth, MCPTransport + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + server = MCPServer( + server_id="github-configured", + name="github_mcp", + url="https://api.githubcopilot.com/mcp/", + transport=MCPTransport.http, + auth_type=MCPAuth.oauth2, + oauth2_flow="authorization_code", + authorization_url=None, + token_url=None, + configured_authorization_url="https://github.com/login/oauth/authorize", + configured_token_url="https://github.com/login/oauth/access_token", + ) + assert _oauth_endpoints_unresolved(server) is False + + def test_passthrough_authorization_code_round_trips_and_rejects_hostile_input(): """The passthrough gateway code seals and recovers the ephemeral DCR client and upstream code, and is total over hostile input: a raw upstream code opens to None, and a tampered or From 8fcbc357e591435176248c995a1ce1cb2e0f4e51 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 12:45:25 -0700 Subject: [PATCH 2/6] fix(mcp): gate deferred oauth discovery on the endpoint each flow needs and read the resolved server The token exchange no longer joins deferred discovery when the token url is already stored, so it cannot 503 over an unreachable issuer it needs nothing from. After a request joins discovery, authorize and token now read the resolved server for the DCR bridge relay decision and the rest of the flow, so a registration endpoint resolved mid-request routes a front-door client to its own redirect binding. The encrypt seam in the issuer-yield authorize test now uses a real salt key instead of patching an SDK internal. --- .../mcp_server/discoverable_endpoints.py | 95 ++++++----- .../mcp_server/test_discoverable_endpoints.py | 150 ++++++++++++++++-- 2 files changed, 191 insertions(+), 54 deletions(-) diff --git a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py index 6ede1c553e8..928373d93d8 100644 --- a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py +++ b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py @@ -3,7 +3,7 @@ import html as _html import json import secrets import time -from collections.abc import Mapping +from collections.abc import Callable, Mapping from datetime import datetime, timezone from typing import TYPE_CHECKING, Any, Final, Literal, Optional from urllib.parse import parse_qsl, urlencode, urlparse, urlunparse @@ -663,15 +663,18 @@ def _endpoint_not_configured_detail( ) -async def _server_with_oauth_endpoints(mcp_server: MCPServer) -> MCPServer: - """Join deferred OAuth discovery only when this server still has no authorize URL. +async def _server_with_oauth_endpoints( + mcp_server: MCPServer, + needed_endpoint: Callable[[MCPServer], str | None], +) -> MCPServer: + """Join deferred OAuth discovery only when the endpoint this caller needs is still missing. Admin-entered endpoints live on ``configured_*`` after an anchored issuer empties the - resolved fields. Those already let authorize/token run, so discovery is not awaited - and cannot 503 over a leftover pin. A server with nothing configured still joins the - deferred task; no slot is a no-op and the caller 400s. + resolved fields. A caller whose needed endpoint already resolves never awaits discovery + and cannot 503 over a leftover pin. A server still missing it joins the deferred task; + no slot is a no-op and the caller 400s. """ - if mcp_server.effective_authorization_url is not None: + if needed_endpoint(mcp_server) is not None: return mcp_server from litellm.proxy._experimental.mcp_server.mcp_server_manager import ( # noqa: PLC0415 # circular import with mcp_server_manager at module load global_mcp_server_manager, @@ -829,7 +832,7 @@ async def authorize_with_server( ephemeral_dcr_client: "EphemeralDcrClient | None" = None, ): _raise_if_not_oauth2(mcp_server) - resolved_server: Final = await _server_with_oauth_endpoints(mcp_server) + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, lambda s: s.effective_authorization_url) if resolved_server.effective_authorization_url is None: raise HTTPException( status_code=400, @@ -841,7 +844,7 @@ async def authorize_with_server( ), ) - if mcp_server.is_dcr_bridge: + if resolved_server.is_dcr_bridge: # Enforce S256 PKCE on both bridge arms. The relay arm forwards the validated, # now-non-optional pair to the upstream authorize; the short-circuit arm keeps # calling this for its enforcement side effect, then falls through to the gateway @@ -850,9 +853,9 @@ async def authorize_with_server( # A gateway-minted ephemeral client is registered against {base}/callback, so its # flow must run the short-circuit arm; the relay arm is only for clients that # registered themselves through the front door and hold their own redirect binding. - if _dcr_bridge_relays_client_registration(mcp_server) and ephemeral_dcr_client is None: + if _dcr_bridge_relays_client_registration(resolved_server) and ephemeral_dcr_client is None: return _redirect_to_upstream_authorize( - mcp_server=mcp_server, + mcp_server=resolved_server, client_id=client_id, redirect_uri=redirect_uri, state=state, @@ -878,7 +881,7 @@ async def authorize_with_server( # litellm key, so the browser session is the only identity source; without one there is nothing to # bind, so send the user through login first. Every other oauth2 server keeps the identity-less state. litellm_user_id: str | None = None - if mcp_server.is_dcr_bridge and mcp_server.is_oauth_delegate: + if resolved_server.is_dcr_bridge and resolved_server.is_oauth_delegate: from litellm.proxy._experimental.mcp_server.byok_oauth_endpoints import ( # noqa: PLC0415 # inline import avoids a module-load circular import _user_id_from_session_cookie, ) @@ -888,7 +891,7 @@ async def authorize_with_server( return _redirect_to_litellm_login(request) denial: Final = await _bridge_authorize_access_denial( litellm_user_id=litellm_user_id, - mcp_server=mcp_server, + mcp_server=resolved_server, redirect_uri=redirect_uri, state=state, ) @@ -902,7 +905,7 @@ async def authorize_with_server( code_challenge_method=code_challenge_method, client_redirect_uri=redirect_uri, litellm_user_id=litellm_user_id, - mcp_server_id=mcp_server.server_id if (litellm_user_id or ephemeral_dcr_client) else None, + mcp_server_id=resolved_server.server_id if (litellm_user_id or ephemeral_dcr_client) else None, dcr_client_id=ephemeral_dcr_client.client_id if ephemeral_dcr_client else None, dcr_client_secret=ephemeral_dcr_client.client_secret if ephemeral_dcr_client else None, dcr_token_endpoint_auth_method=ephemeral_dcr_client.token_endpoint_auth_method @@ -912,22 +915,22 @@ async def authorize_with_server( relay_state: Final = secrets.token_urlsafe(_OAUTH_STATE_HANDLE_BYTES) params: Final = { - "client_id": mcp_server.client_id if mcp_server.client_id else client_id, + "client_id": resolved_server.client_id if resolved_server.client_id else client_id, "redirect_uri": f"{request_base_url}/callback", "state": relay_state, "response_type": response_type or "code", } if scope: params["scope"] = scope - elif mcp_server.scopes: - params["scope"] = " ".join(mcp_server.scopes) + elif resolved_server.scopes: + params["scope"] = " ".join(resolved_server.scopes) if code_challenge: params["code_challenge"] = code_challenge if code_challenge_method: params["code_challenge_method"] = code_challenge_method - upstream_resource: Final = resolve_upstream_resource(mcp_server) + upstream_resource: Final = resolve_upstream_resource(resolved_server) if upstream_resource: params["resource"] = upstream_resource @@ -964,7 +967,7 @@ async def exchange_token_with_server( if grant_type not in ("authorization_code", "refresh_token"): raise HTTPException(status_code=400, detail="Unsupported grant_type") - resolved_server: Final = await _server_with_oauth_endpoints(mcp_server) + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, lambda s: s.effective_token_url) token_url: Final = resolved_server.effective_token_url if token_url is None: raise HTTPException( @@ -985,16 +988,16 @@ async def exchange_token_with_server( # recovered from a sealed code) must authenticate the way its own registration was granted, # not the way the server row is configured; callers that carry no method keep the row's method # as before. - resolved_client_id: Final = mcp_server.client_id if mcp_server.client_id else client_id - resolved_client_secret: Final = mcp_server.client_secret if mcp_server.client_id else client_secret + resolved_client_id: Final = resolved_server.client_id if resolved_server.client_id else client_id + resolved_client_secret: Final = resolved_server.client_secret if resolved_server.client_id else client_secret resolved_auth_method: Final = ( - mcp_server.token_endpoint_auth_method - if mcp_server.client_id - else (client_token_endpoint_auth_method or mcp_server.token_endpoint_auth_method) + resolved_server.token_endpoint_auth_method + if resolved_server.client_id + else (client_token_endpoint_auth_method or resolved_server.token_endpoint_auth_method) ) try: token_request: Final = build_upstream_oauth2_token_request( - mcp_server, + resolved_server, auth_method=resolved_auth_method, client_id=resolved_client_id, client_secret=resolved_client_secret, @@ -1007,14 +1010,14 @@ async def exchange_token_with_server( bridge_upstream_refresh: SecretStr | None = None bridge_upstream_scope: str | None = None refresh_request_scope: str | None = None - is_bridge: Final = mcp_server.is_oauth_delegate and mcp_server.is_dcr_bridge + is_bridge: Final = resolved_server.is_oauth_delegate and resolved_server.is_dcr_bridge if grant_type == "refresh_token": # Phase 1 for a bridge refresh: open the client's refresh envelope, re-validate the sealed # identity, and unwrap the real upstream refresh token BEFORE building token_data, so the exchange # sends the upstream token and never the envelope. A failure returns without touching the upstream. if is_bridge: - prepared_refresh: Final = await _prepare_bridge_refresh(mcp_server, refresh_token) + prepared_refresh: Final = await _prepare_bridge_refresh(resolved_server, refresh_token) if not isinstance(prepared_refresh, _BridgeRefreshReady): return _bridge_mint_error_response(prepared_refresh) bridge_mint_ready = prepared_refresh.ready @@ -1051,13 +1054,13 @@ async def exchange_token_with_server( # A raw upstream code (scripted path) opens to None and the code is used as-is. bridge_identity = open_bridge_authorization_code(code) if bridge_identity is not None: - if bridge_identity.mcp_server_id != mcp_server.server_id: + if bridge_identity.mcp_server_id != resolved_server.server_id: raise HTTPException( status_code=400, detail="Authorization code was issued for a different MCP server", ) code = bridge_identity.upstream_code - bridge_token_relay: Final = _dcr_bridge_relays_client_registration(mcp_server) + bridge_token_relay: Final = _dcr_bridge_relays_client_registration(resolved_server) if bridge_token_relay and not redirect_uri: raise HTTPException( status_code=400, @@ -1079,7 +1082,7 @@ async def exchange_token_with_server( # Phase 1 for a bridge authorization_code mint: resolve identity (the SSO user recovered above, or # the presented litellm key) and the envelope keys BEFORE the exchange consumes the single-use code. if is_bridge: - prepared: Final = await _prepare_bridge_mint(request, mcp_server, bridge_identity) + prepared: Final = await _prepare_bridge_mint(request, resolved_server, bridge_identity) if not isinstance(prepared, _BridgeMintReady): return _bridge_mint_error_response(prepared) bridge_mint_ready = prepared @@ -1096,8 +1099,8 @@ async def exchange_token_with_server( except httpx.HTTPStatusError as exc: fault: Final = classify_upstream_token_rejection( exc.response, - credential_source=_token_credential_source(mcp_server), - log_context=mcp_server.server_id, + credential_source=_token_credential_source(resolved_server), + log_context=resolved_server.server_id, ) upstream_rejected_bridge_refresh: Final = ( is_bridge @@ -1110,7 +1113,7 @@ async def exchange_token_with_server( "bridge refresh: the upstream rejected the sealed refresh token for server=%s with " "invalid_grant (revoked or expired at the IdP); returning invalid_grant so the client " "re-runs authorization_code rather than an opaque upstream error", - mcp_server.server_id, + resolved_server.server_id, ) return _bridge_mint_error_response("invalid_refresh") return render_token_fault(fault) @@ -1123,22 +1126,22 @@ async def exchange_token_with_server( # Validate token response against server-configured rules before any storage. # This rejects tokens from wrong Slack workspaces, Atlassian orgs, etc. - if mcp_server.token_validation and isinstance(mcp_server.token_validation, dict): + if resolved_server.token_validation and isinstance(resolved_server.token_validation, dict): _validate_token_response( token_response=token_response, - validation_rules=mcp_server.token_validation, - server_id=mcp_server.server_id, + validation_rules=resolved_server.token_validation, + server_id=resolved_server.server_id, ) # Store server-side when the server is configured for per-user OAuth and # the calling client has provided a valid LiteLLM identity. # Errors are non-fatal: the token is still returned to the client. - if mcp_server.needs_user_oauth_token: + if resolved_server.needs_user_oauth_token: user_id: Final = await _extract_user_id_from_request(request) if user_id: try: await _store_per_user_token_server_side( - server=mcp_server, + server=resolved_server, user_id=user_id, token_response=token_response, ) @@ -1146,7 +1149,7 @@ async def exchange_token_with_server( verbose_logger.warning( "exchange_token_with_server: server-side storage failed for user=%s server=%s: %s", user_id, - mcp_server.server_id, + resolved_server.server_id, exc, ) else: @@ -1156,7 +1159,7 @@ async def exchange_token_with_server( "requires the stored token, so the client will be challenged with 401 on reconnect. " "Ensure the request carries a valid LiteLLM key (x-litellm-api-key or Authorization), " "or store it via POST /mcp/server/{id}/oauth-user-credential.", - mcp_server.server_id, + resolved_server.server_id, ) # A DCR-bridge oauth_delegate server hands the client a gateway-bound envelope (identity plus the @@ -1167,7 +1170,9 @@ async def exchange_token_with_server( token_response = {**token_response, "scope": refresh_request_scope} # Phase 3: seal the upstream grant into the client-held envelope; failures map through the same # OAuth-shaped response as the phase-1 preconditions. - minted: Final = _finish_bridge_mint(bridge_mint_ready, mcp_server, token_response, datetime.now(timezone.utc)) + minted: Final = _finish_bridge_mint( + bridge_mint_ready, resolved_server, token_response, datetime.now(timezone.utc) + ) return minted if isinstance(minted, JSONResponse) else _bridge_mint_error_response(minted) raw_access_token: Final = token_response.get("access_token") if isinstance(token_response, dict) else None @@ -1682,7 +1687,7 @@ async def register_client_with_server( ): return dummy_return - resolved_server: Final = await _server_with_oauth_endpoints(mcp_server) + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, lambda s: s.effective_authorization_url) if resolved_server.effective_authorization_url is None: raise HTTPException( status_code=400, @@ -1715,13 +1720,15 @@ async def register_client_with_server( response: Final = await _post_dcr_registration( registration_url=registration_url, register_data=register_data, - server_id=mcp_server.server_id, + server_id=resolved_server.server_id, ) token_response = response.json() if persist_credentials and not bridge_relay: - persistence_result = await _persist_dcr_client_registration(mcp_server, token_response, current_redirect_uri) + persistence_result = await _persist_dcr_client_registration( + resolved_server, token_response, current_redirect_uri + ) if persistence_result == "reused": return dummy_return diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py index 790b5c7ad22..46c9c21e80d 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py @@ -8855,7 +8855,7 @@ async def test_authorize_wall_names_the_issuer_for_anchored_servers(): @pytest.mark.asyncio -async def test_authorize_uses_admin_entered_github_oauth_urls_after_issuer_yield(): +async def test_authorize_uses_admin_entered_github_oauth_urls_after_issuer_yield(monkeypatch): """GitHub MCP servers store Authorization URL and Token URL on the row. 1.99 can empty the resolved authorization_url when a leftover issuer is treated as a pin (RFC 8414 yield). The UI authorize must still redirect to the admin-entered GitHub authorize URL @@ -8887,15 +8887,14 @@ async def test_authorize_uses_admin_entered_github_oauth_urls_after_issuer_yield mock_request.base_url = "https://litellm.example.com/" mock_request.headers = {} - with patch("litellm.proxy._experimental.mcp_server.discoverable_endpoints.encrypt_value_helper") as mock_encrypt: - mock_encrypt.return_value = "mocked_encrypted_state" - response = await authorize_with_server( - request=mock_request, - mcp_server=server, - client_id="github-app-client", - redirect_uri="http://127.0.0.1:60108/callback", - state="state123", - ) + monkeypatch.setenv("LITELLM_SALT_KEY", "sk-test-salt-for-lit-6255") + response = await authorize_with_server( + request=mock_request, + mcp_server=server, + client_id="github-app-client", + redirect_uri="http://127.0.0.1:60108/callback", + state="state123", + ) assert response.status_code == 307 assert "https://github.com/login/oauth/authorize" in response.headers["location"] @@ -8926,6 +8925,137 @@ def test_oauth_endpoints_count_admin_entered_urls_as_resolved(): assert _oauth_endpoints_unresolved(server) is False +@pytest.mark.asyncio +async def test_token_exchange_with_configured_token_url_never_joins_discovery(monkeypatch): + """A server can hold an admin-entered Token URL while its Authorization URL is absent. The + token exchange must post to that stored endpoint without awaiting deferred discovery, which + can 503 against an unreachable issuer even though nothing it resolves is needed here.""" + from litellm.proxy._experimental.mcp_server import ( + discoverable_endpoints, + mcp_server_manager, + ) + from litellm.proxy._experimental.mcp_server.discoverable_endpoints import ( + exchange_token_with_server, + ) + from litellm.types.mcp import MCPAuth, MCPTransport + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + server = MCPServer( + server_id="token-url-only", + name="token_url_only", + server_name="token_url_only", + alias="token_url_only", + url="https://mcp.example.com/mcp/", + transport=MCPTransport.http, + auth_type=MCPAuth.oauth2, + client_id="cid", + client_secret="cs", + authorization_url=None, + token_url=None, + issuer="https://idp.example.com", + issuer_is_anchored=True, + configured_token_url="https://idp.example.com/oauth/token", + ) + + async def fail_discovery(_srv): + raise AssertionError("the exchange joined deferred discovery despite a stored token url") + + monkeypatch.setattr( + mcp_server_manager.global_mcp_server_manager, + "ensure_oauth_metadata_discovered", + fail_discovery, + ) + fake_http_response = MagicMock() + fake_http_response.json.return_value = {"access_token": "tok", "token_type": "Bearer"} + fake_http_response.raise_for_status = MagicMock() + fake_http_client = MagicMock() + fake_http_client.post = AsyncMock(return_value=fake_http_response) + monkeypatch.setattr( + discoverable_endpoints, + "get_async_httpx_client", + lambda llm_provider: fake_http_client, + ) + mock_request = MagicMock() + mock_request.base_url = "https://litellm.example.com/" + mock_request.headers = {} + + response = await exchange_token_with_server( + request=mock_request, + mcp_server=server, + grant_type="authorization_code", + code="upstream-code", + redirect_uri="http://127.0.0.1:3000/cb", + client_id="cid", + client_secret=None, + code_verifier=None, + ) + + assert response.status_code == 200 + assert fake_http_client.post.await_args.args[0] == "https://idp.example.com/oauth/token" + + +@pytest.mark.asyncio +async def test_bridge_authorize_relays_with_registration_url_resolved_by_deferred_discovery(monkeypatch): + """When deferred discovery resolves a DCR-bridge server during the authorize request, the + relay-vs-short-circuit call must read the resolved server: a client that registered itself + through the front door keeps its own redirect binding instead of being routed through the + gateway callback the upstream never granted it.""" + from litellm.proxy._experimental.mcp_server import mcp_server_manager + from litellm.proxy._experimental.mcp_server.discoverable_endpoints import ( + authorize_with_server, + ) + from litellm.types.mcp import MCPAuth, MCPTransport + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + server = MCPServer( + server_id="bridge-deferred", + name="bridge_deferred", + server_name="bridge_deferred", + alias="bridge_deferred", + url="https://mcp.example.com/mcp/", + transport=MCPTransport.http, + auth_type=MCPAuth.true_passthrough, + dcr_bridge=True, + authorization_url=None, + token_url=None, + registration_url=None, + ) + resolved = server.model_copy( + update={ + "authorization_url": "https://idp.example.com/oauth/authorize", + "token_url": "https://idp.example.com/oauth/token", + "registration_url": "https://idp.example.com/oauth/register", + } + ) + + async def resolve_discovery(_srv): + return resolved + + monkeypatch.setattr( + mcp_server_manager.global_mcp_server_manager, + "ensure_oauth_metadata_discovered", + resolve_discovery, + ) + mock_request = MagicMock() + mock_request.base_url = "https://litellm.example.com/" + mock_request.headers = {} + + response = await authorize_with_server( + request=mock_request, + mcp_server=server, + client_id="front-door-client", + redirect_uri="http://127.0.0.1:60110/client-callback", + state="state456", + code_challenge="E9Melhoa2OwvFrEMTJguCHaoeK1t8URWbuGJSstw-cM", + code_challenge_method="S256", + ) + + assert response.status_code == 307 + location = response.headers["location"] + assert location.startswith("https://idp.example.com/oauth/authorize") + assert "redirect_uri=http%3A%2F%2F127.0.0.1%3A60110%2Fclient-callback" in location + + def test_passthrough_authorization_code_round_trips_and_rejects_hostile_input(): """The passthrough gateway code seals and recovers the ephemeral DCR client and upstream code, and is total over hostile input: a raw upstream code opens to None, and a tampered or From 19e6f03a3c08c71c9307d5588d7277604acb4c64 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 13:39:35 -0700 Subject: [PATCH 3/6] fix(mcp): let root oauth routes defer discovery to the endpoint-gated flow join --- .../mcp_server/discoverable_endpoints.py | 28 +--- .../mcp_server/mcp_server_manager.py | 8 -- .../mcp_server/test_discoverable_endpoints.py | 125 ++++++++++++++---- 3 files changed, 108 insertions(+), 53 deletions(-) diff --git a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py index 928373d93d8..6274cfcef8b 100644 --- a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py +++ b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py @@ -1785,17 +1785,10 @@ async def authorize( lookup_name: Final[str | None] = mcp_server_name or client_id client_ip: Final = IPAddressUtils.get_mcp_client_ip(request) mcp_server = ( - await global_mcp_server_manager.get_resolved_mcp_server_by_name(lookup_name, client_ip=client_ip) - if lookup_name - else None + global_mcp_server_manager.get_mcp_server_by_name(lookup_name, client_ip=client_ip) if lookup_name else None ) if mcp_server is None and mcp_server_name is None: - unresolved_server: Final = _resolve_oauth2_server_for_root_endpoints(client_ip=client_ip) - mcp_server = ( - await global_mcp_server_manager.ensure_oauth_metadata_discovered(unresolved_server) - if unresolved_server is not None - else None - ) + mcp_server = _resolve_oauth2_server_for_root_endpoints(client_ip=client_ip) if mcp_server is None: raise HTTPException(status_code=404, detail="MCP server not found") _raise_if_not_oauth2(mcp_server) @@ -1876,14 +1869,9 @@ async def token_endpoint( lookup_name: Final = mcp_server_name or client_id client_ip: Final = IPAddressUtils.get_mcp_client_ip(request) - mcp_server = await global_mcp_server_manager.get_resolved_mcp_server_by_name(lookup_name, client_ip=client_ip) + mcp_server = global_mcp_server_manager.get_mcp_server_by_name(lookup_name, client_ip=client_ip) if mcp_server is None and mcp_server_name is None: - unresolved_server: Final = _resolve_oauth2_server_for_root_endpoints(client_ip=client_ip) - mcp_server = ( - await global_mcp_server_manager.ensure_oauth_metadata_discovered(unresolved_server) - if unresolved_server is not None - else None - ) + mcp_server = _resolve_oauth2_server_for_root_endpoints(client_ip=client_ip) if mcp_server is None: raise HTTPException(status_code=404, detail="MCP server not found") return await exchange_token_with_server( @@ -2714,10 +2702,9 @@ async def register_client(request: Request, mcp_server_name: str | None = None): return await register_aggregate_client(request=request, request_body=data) resolved: Final = _resolve_oauth2_server_for_root_endpoints(client_ip=client_ip) if resolved: - resolved_server: Final = await global_mcp_server_manager.ensure_oauth_metadata_discovered(resolved) return await register_client_with_server( request=request, - mcp_server=resolved_server, + mcp_server=resolved, client_name=data.get("client_name", ""), grant_types=data.get("grant_types", []), response_types=data.get("response_types", []), @@ -2727,10 +2714,7 @@ async def register_client(request: Request, mcp_server_name: str | None = None): ) return dummy_return - mcp_server: Final = await global_mcp_server_manager.get_resolved_mcp_server_by_name( - mcp_server_name, - client_ip=client_ip, - ) + mcp_server: Final = global_mcp_server_manager.get_mcp_server_by_name(mcp_server_name, client_ip=client_ip) if mcp_server is None: return dummy_return return await register_client_with_server( diff --git a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py index fe0c73a5efc..308813039ca 100644 --- a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py +++ b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py @@ -6205,14 +6205,6 @@ class MCPServerManager: return server return None - async def get_resolved_mcp_server_by_name( - self, - server_name: str, - client_ip: str | None = None, - ) -> MCPServer | None: - server: Final = self.get_mcp_server_by_name(server_name, client_ip=client_ip) - return await self.ensure_oauth_metadata_discovered(server) if server is not None else None - def get_filtered_registry(self, client_ip: str | None = None) -> dict[str, MCPServer]: """ Get registry filtered by client IP access control. diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py index 46c9c21e80d..f67e9a67d53 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py @@ -79,24 +79,23 @@ def _resolved_oauth_metadata(): @pytest.mark.asyncio -async def test_authorize_resolves_cold_oauth_metadata(): +async def test_authorize_resolves_cold_oauth_metadata(monkeypatch): + """The route hands the registered server to the flow, whose deferred-discovery join resolves + the cold metadata; the redirect must land on the discovered authorization endpoint.""" from litellm.proxy._experimental.mcp_server import discoverable_endpoints from litellm.proxy._experimental.mcp_server.mcp_server_manager import global_mcp_server_manager + monkeypatch.setenv("LITELLM_SALT_KEY", "sk-test-salt-for-lit-6255") server = _unresolved_oauth_server() global_mcp_server_manager.registry[server.server_id] = server global_mcp_server_manager._set_oauth_discovery_deferred(server.server_id, True) request = _mock_callback_request("https://litellm.example.com/") - expected = MagicMock() - with ( - patch.object( - global_mcp_server_manager, - "_discover_oauth_metadata_for_server", - new=AsyncMock(return_value=_resolved_oauth_metadata()), - ) as discovery, - patch.object(discoverable_endpoints, "authorize_with_server", new=AsyncMock(return_value=expected)) as relay, - ): + with patch.object( + global_mcp_server_manager, + "_discover_oauth_metadata_for_server", + new=AsyncMock(return_value=_resolved_oauth_metadata()), + ) as discovery: response = await discoverable_endpoints.authorize( request=request, client_id="client-id", @@ -105,12 +104,14 @@ async def test_authorize_resolves_cold_oauth_metadata(): ) discovery.assert_awaited_once_with(server) - assert relay.await_args.kwargs["mcp_server"].authorization_url == "https://idp.example.com/authorize" - assert response is expected + assert response.status_code == 307 + assert response.headers["location"].startswith("https://idp.example.com/authorize") @pytest.mark.asyncio async def test_token_resolves_cold_oauth_metadata(): + """The route hands the registered server to the exchange, whose deferred-discovery join + resolves the cold metadata; the exchange must post to the discovered token endpoint.""" from litellm.proxy._experimental.mcp_server import discoverable_endpoints from litellm.proxy._experimental.mcp_server.mcp_server_manager import global_mcp_server_manager @@ -118,7 +119,11 @@ async def test_token_resolves_cold_oauth_metadata(): global_mcp_server_manager.registry[server.server_id] = server global_mcp_server_manager._set_oauth_discovery_deferred(server.server_id, True) request = _mock_callback_request("https://litellm.example.com/") - expected = MagicMock() + fake_http_response = MagicMock() + fake_http_response.json.return_value = {"access_token": "tok", "token_type": "Bearer"} + fake_http_response.raise_for_status = MagicMock() + fake_http_client = MagicMock() + fake_http_client.post = AsyncMock(return_value=fake_http_response) with ( patch.object( @@ -127,8 +132,10 @@ async def test_token_resolves_cold_oauth_metadata(): new=AsyncMock(return_value=_resolved_oauth_metadata()), ) as discovery, patch.object( - discoverable_endpoints, "exchange_token_with_server", new=AsyncMock(return_value=expected) - ) as relay, + discoverable_endpoints, + "get_async_httpx_client", + new=lambda llm_provider: fake_http_client, + ), ): response = await discoverable_endpoints.token_endpoint( request=request, @@ -139,20 +146,26 @@ async def test_token_resolves_cold_oauth_metadata(): ) discovery.assert_awaited_once_with(server) - assert relay.await_args.kwargs["mcp_server"].token_url == "https://idp.example.com/token" - assert response is expected + assert response.status_code == 200 + assert fake_http_client.post.await_args.args[0] == "https://idp.example.com/token" @pytest.mark.asyncio async def test_register_resolves_cold_oauth_metadata(): + """The route hands the registered server to the registration flow, whose deferred-discovery + join resolves the cold metadata; DCR must post to the discovered registration endpoint.""" from litellm.proxy._experimental.mcp_server import discoverable_endpoints from litellm.proxy._experimental.mcp_server.mcp_server_manager import global_mcp_server_manager - server = _unresolved_oauth_server() + server = _unresolved_oauth_server().model_copy(update={"client_id": None}) global_mcp_server_manager.registry[server.server_id] = server global_mcp_server_manager._set_oauth_discovery_deferred(server.server_id, True) request = _mock_callback_request("https://litellm.example.com/") - expected = MagicMock() + fake_http_response = MagicMock() + fake_http_response.json.return_value = {"client_id": "generated-client", "client_secret": "generated-secret"} + fake_http_response.raise_for_status = MagicMock() + fake_http_client = MagicMock() + fake_http_client.post = AsyncMock(return_value=fake_http_response) with ( patch.object( @@ -162,14 +175,16 @@ async def test_register_resolves_cold_oauth_metadata(): ) as discovery, patch.object(discoverable_endpoints, "_read_request_body", new=AsyncMock(return_value={})), patch.object( - discoverable_endpoints, "register_client_with_server", new=AsyncMock(return_value=expected) - ) as relay, + discoverable_endpoints, + "get_async_httpx_client", + new=lambda llm_provider: fake_http_client, + ), ): response = await discoverable_endpoints.register_client(request=request, mcp_server_name=server.server_name) discovery.assert_awaited_once_with(server) - assert relay.await_args.kwargs["mcp_server"].registration_url == "https://idp.example.com/register" - assert response is expected + assert response.status_code == 200 + assert fake_http_client.post.await_args.args[0] == "https://idp.example.com/register" @pytest.fixture @@ -8994,6 +9009,70 @@ async def test_token_exchange_with_configured_token_url_never_joins_discovery(mo assert fake_http_client.post.await_args.args[0] == "https://idp.example.com/oauth/token" +@pytest.mark.asyncio +async def test_root_token_route_with_configured_token_url_never_joins_discovery(monkeypatch): + """A root POST /token that falls back to the sole OAuth2 server must reach the exchange's + endpoint-gated discovery join instead of awaiting full discovery at the route: with the + token url admin-entered, a failing or slow discovery must not turn the exchange into a 503.""" + from litellm.proxy._experimental.mcp_server import ( + discoverable_endpoints, + mcp_server_manager, + ) + from litellm.types.mcp import MCPAuth, MCPTransport + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + manager = mcp_server_manager.global_mcp_server_manager + server = MCPServer( + server_id="sole-token-url-only", + name="sole_token_url_only", + server_name="sole_token_url_only", + alias="sole_token_url_only", + url="https://mcp.example.com/mcp/", + transport=MCPTransport.http, + auth_type=MCPAuth.oauth2, + client_id="cid", + client_secret="cs", + issuer="https://idp.example.com", + issuer_is_anchored=True, + configured_token_url="https://idp.example.com/oauth/token", + ) + saved_registry = dict(manager.registry) + manager.registry.clear() + manager.registry[server.server_id] = server + manager._set_oauth_discovery_deferred(server.server_id, True) + + async def fail_discovery(_srv): + raise AssertionError("the root token route joined deferred discovery despite a stored token url") + + monkeypatch.setattr(manager, "ensure_oauth_metadata_discovered", fail_discovery) + fake_http_response = MagicMock() + fake_http_response.json.return_value = {"access_token": "tok", "token_type": "Bearer"} + fake_http_response.raise_for_status = MagicMock() + fake_http_client = MagicMock() + fake_http_client.post = AsyncMock(return_value=fake_http_response) + monkeypatch.setattr( + discoverable_endpoints, + "get_async_httpx_client", + lambda llm_provider: fake_http_client, + ) + request = _mock_callback_request("https://litellm.example.com/") + + try: + response = await discoverable_endpoints.token_endpoint( + request=request, + grant_type="authorization_code", + code="upstream-code", + redirect_uri="http://127.0.0.1:3000/cb", + client_id="unregistered-dcr-client", + ) + finally: + manager.registry.clear() + manager.registry.update(saved_registry) + + assert response.status_code == 200 + assert fake_http_client.post.await_args.args[0] == "https://idp.example.com/oauth/token" + + @pytest.mark.asyncio async def test_bridge_authorize_relays_with_registration_url_resolved_by_deferred_discovery(monkeypatch): """When deferred discovery resolves a DCR-bridge server during the authorize request, the From 465ebb1bdd3a143c72179b2babe027d51ebb12ba Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 14:08:09 -0700 Subject: [PATCH 4/6] fix(mcp): join discovery for a clientless DCR bridge still missing its registration endpoint --- .../mcp_server/discoverable_endpoints.py | 15 ++++- .../mcp_server/test_discoverable_endpoints.py | 60 +++++++++++++++++++ 2 files changed, 73 insertions(+), 2 deletions(-) diff --git a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py index 6274cfcef8b..46feeab4dc3 100644 --- a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py +++ b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py @@ -832,7 +832,7 @@ async def authorize_with_server( ephemeral_dcr_client: "EphemeralDcrClient | None" = None, ): _raise_if_not_oauth2(mcp_server) - resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, lambda s: s.effective_authorization_url) + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, _register_flow_needed_endpoint) if resolved_server.effective_authorization_url is None: raise HTTPException( status_code=400, @@ -1653,6 +1653,17 @@ async def resolve_ephemeral_dcr_client( return await mint_ephemeral_dcr_client(request, mcp_server) +def _register_flow_needed_endpoint(mcp_server: MCPServer) -> str | None: + """The register flow's deferred-discovery join gate. A DCR bridge with no admin-configured + client can only register callers through the upstream's registration endpoint + (``_oauth_endpoints_unresolved`` keeps its discovery slot armed for exactly this shape), so + the flow must keep joining discovery while registration is still missing instead of silently + degrading to the dummy short-circuit. Every other shape only needs the authorization url.""" + if mcp_server.is_dcr_bridge and not mcp_server.client_id and mcp_server.effective_registration_url is None: + return None + return mcp_server.effective_authorization_url + + async def register_client_with_server( request: Request, mcp_server: MCPServer, @@ -1687,7 +1698,7 @@ async def register_client_with_server( ): return dummy_return - resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, lambda s: s.effective_authorization_url) + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, _register_flow_needed_endpoint) if resolved_server.effective_authorization_url is None: raise HTTPException( status_code=400, diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py index f67e9a67d53..b1a99b498e5 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py @@ -187,6 +187,66 @@ async def test_register_resolves_cold_oauth_metadata(): assert fake_http_client.post.await_args.args[0] == "https://idp.example.com/register" +@pytest.mark.asyncio +async def test_register_route_bridge_missing_registration_url_joins_discovery(): + """A clientless DCR bridge whose authorize and token urls are admin-entered still relays + registration upstream: the flow must join deferred discovery for the missing registration + endpoint instead of short-circuiting to dummy credentials because authorization resolves.""" + import json + + from litellm.proxy._experimental.mcp_server import discoverable_endpoints + from litellm.proxy._experimental.mcp_server.mcp_server_manager import global_mcp_server_manager + from litellm.proxy._types import MCPTransport + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + server = MCPServer( + server_id="bridge-partial-metadata", + name="bridge_partial_metadata", + server_name="bridge_partial_metadata", + alias="bridge_partial_metadata", + url="https://mcp.example.com/mcp", + transport=MCPTransport.http, + auth_type=MCPAuth.oauth_delegate, + dcr_bridge=True, + client_id=None, + authorization_url="https://idp.example.com/authorize", + token_url="https://idp.example.com/token", + ) + global_mcp_server_manager.registry[server.server_id] = server + global_mcp_server_manager._set_oauth_discovery_deferred(server.server_id, True) + request = _mock_callback_request("https://litellm.example.com/") + fake_http_response = MagicMock() + fake_http_response.json.return_value = {"client_id": "generated-client", "client_secret": "generated-secret"} + fake_http_response.raise_for_status = MagicMock() + fake_http_client = MagicMock() + fake_http_client.post = AsyncMock(return_value=fake_http_response) + + with ( + patch.object( # test-quality-ok: innermost discovery seam on a module-global manager; the route-to-flow join under test stays real + global_mcp_server_manager, + "_discover_oauth_metadata_for_server", + new=AsyncMock(return_value=_resolved_oauth_metadata()), + ) as discovery, + patch.object( # test-quality-ok: the MagicMock Request carries no body; this seam feeds the RFC 7591 redirect_uris + discoverable_endpoints, + "_read_request_body", + new=AsyncMock(return_value={"redirect_uris": ["https://client.example.com/cb"]}), + ), + patch.object( # test-quality-ok: keeps the DCR POST off the network so its target URL can be asserted + discoverable_endpoints, + "get_async_httpx_client", + new=lambda llm_provider: fake_http_client, + ), + ): + response = await discoverable_endpoints.register_client(request=request, mcp_server_name=server.server_name) + + discovery.assert_awaited_once_with(server) + assert fake_http_client.post.await_args.args[0] == "https://idp.example.com/register" + assert fake_http_client.post.await_args.kwargs["json"]["redirect_uris"] == ["https://client.example.com/cb"] + assert response.status_code == 200 + assert json.loads(response.body.decode("utf-8"))["client_id"] == "generated-client" + + @pytest.fixture def trust_xff(): """Force ``IPAddressUtils.is_request_from_trusted_proxy`` to True. From ce8d6f7d25ae52cf0ae1ea38d76cee94233ca982 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 15:31:45 -0700 Subject: [PATCH 5/6] fix(mcp): token refresh and M2M egress honor the admin-entered token URL --- litellm/proxy/_experimental/mcp_server/db.py | 4 +-- .../mcp_server/oauth2_token_cache.py | 13 +++++----- .../outbound_credentials/adapter.py | 4 +-- .../authz_code_refresher.py | 7 ++++-- .../outbound_credentials/test_adapter.py | 19 ++++++++++++++ .../test_authz_code_refresher.py | 25 +++++++++++++++++++ .../mcp_server/test_db_credentials.py | 25 +++++++++++++++++++ .../mcp_server/test_oauth2_token_cache.py | 19 ++++++++++++++ 8 files changed, 104 insertions(+), 12 deletions(-) diff --git a/litellm/proxy/_experimental/mcp_server/db.py b/litellm/proxy/_experimental/mcp_server/db.py index 4aa08020527..cf74cbd187e 100644 --- a/litellm/proxy/_experimental/mcp_server/db.py +++ b/litellm/proxy/_experimental/mcp_server/db.py @@ -1600,7 +1600,7 @@ async def refresh_user_oauth_token( ) -> OAuthCredentialPayload | None: """Attempt to refresh a per-user OAuth2 token using its stored refresh_token. - POSTs to ``server.token_url`` with ``grant_type=refresh_token``. + POSTs to ``server.effective_token_url`` with ``grant_type=refresh_token``. On success: persists the new credential via ``store_user_oauth_credential`` and returns the updated payload dict. @@ -1609,7 +1609,7 @@ async def refresh_user_oauth_token( stale credential and triggering re-authentication. """ refresh_token: Final[str | None] = cred.get("refresh_token") - token_url: Final[str | None] = getattr(server, "token_url", None) + token_url: Final[str | None] = getattr(server, "effective_token_url", None) or getattr(server, "token_url", None) server_id: Final[str] = getattr(server, "server_id", "") client_id: Final[str | None] = getattr(server, "client_id", None) client_secret: Final[str | None] = getattr(server, "client_secret", None) diff --git a/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py b/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py index c76c933c5b5..b3f1da51074 100644 --- a/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py +++ b/litellm/proxy/_experimental/mcp_server/oauth2_token_cache.py @@ -67,7 +67,7 @@ class MCPOAuth2TokenCache(InMemoryCache): rest of the identity rather than stored in a key.""" material: Final = "\x00".join( ( - server.token_url or "", + server.effective_token_url or "", server.client_id or "", server.client_secret or "", " ".join(server.scopes or ()), @@ -82,7 +82,7 @@ class MCPOAuth2TokenCache(InMemoryCache): @staticmethod def _has_client_credentials_config(server: "MCPServer") -> bool: - return bool(server.client_id and server.client_secret and server.token_url) + return bool(server.client_id and server.client_secret and server.effective_token_url) async def async_get_token(self, server: "MCPServer") -> str | None: """Return a valid access token, fetching or refreshing as needed. @@ -112,19 +112,20 @@ class MCPOAuth2TokenCache(InMemoryCache): return token async def _fetch_token(self, server: "MCPServer") -> tuple[str, int]: - """POST to ``token_url`` with ``grant_type=client_credentials``. + """POST to ``effective_token_url`` with ``grant_type=client_credentials``. Returns ``(access_token, ttl_seconds)`` where ttl accounts for the expiry buffer so the cache entry expires before the real token does. """ client: Final = get_async_httpx_client(llm_provider=httpxSpecialProvider.MCP) - if not server.client_id or not server.client_secret or not server.token_url: + token_url: Final = server.effective_token_url + if not server.client_id or not server.client_secret or not token_url: raise ValueError( f"MCP server '{server.server_id}' missing required OAuth2 fields: " f"client_id={bool(server.client_id)}, " f"client_secret={bool(server.client_secret)}, " - f"token_url={bool(server.token_url)}" + f"token_url={bool(token_url)}" ) token_request: Final = build_upstream_oauth2_token_request( @@ -146,7 +147,7 @@ class MCPOAuth2TokenCache(InMemoryCache): ) try: - response: Final = await client.post(server.token_url, data=data, headers=token_request.headers or None) + response: Final = await client.post(token_url, data=data, headers=token_request.headers or None) response.raise_for_status() except httpx.HTTPStatusError as exc: raise ValueError( diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py index be8ec1b8eb3..98e239b1d1d 100644 --- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py +++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py @@ -142,7 +142,7 @@ def _client_credentials_spec(server: MCPServer, resource: str) -> ServerSpec: config=ClientCredentialsConfig( client_id=server.client_id, client_secret=SecretStr(server.client_secret) if server.client_secret else None, - token_url=server.token_url, + token_url=server.effective_token_url, scopes=tuple(server.scopes or ()), audience=server.audience, upstream_resource=resolve_upstream_resource(server), @@ -163,7 +163,7 @@ def _token_exchange_spec(server: MCPServer, resource: str) -> ServerSpec | None: normalizes to ``rfc8693`` so a bad config value cannot crash spec-building. ``audience`` is forwarded only when the operator set it; a missing one is omitted, not derived. """ - endpoint: Final = server.token_exchange_endpoint or server.token_url + endpoint: Final = server.token_exchange_endpoint or server.effective_token_url if not server.client_id or not server.client_secret: return None profile: Final[Literal["rfc8693", "entra_obo"]] = ( diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/authz_code_refresher.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/authz_code_refresher.py index 6ea5756d43d..92bd30694af 100644 --- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/authz_code_refresher.py +++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/authz_code_refresher.py @@ -88,7 +88,10 @@ class AuthorizationCodeRefresher: if token.refresh_token is None: return None server: Final = self._server_lookup(server_id) - if server is None or not server.token_url: + if server is None: + return None + token_url: Final = server.effective_token_url + if not token_url: return None try: @@ -106,7 +109,7 @@ class AuthorizationCodeRefresher: "refresh_token": token.refresh_token, **token_request.body, } - body: Final = await self._token_endpoint(server.token_url, form, token_request.headers) + body: Final = await self._token_endpoint(token_url, form, token_request.headers) if body is None: return None access_token: Final = body.get("access_token") diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py index e336bdc80c2..0020dbf8d61 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py @@ -579,3 +579,22 @@ def test_id_jag_honors_explicit_subject_token_type(): def test_id_jag_half_configured_defers_to_v1(server): # A half-configured server must defer (None) rather than 500 at IdJagConfig construction. assert to_server_spec(server) is None + + +def test_client_credentials_uses_admin_entered_token_url_when_issuer_yield_empties_resolved(): + """A pinned issuer empties the resolved token_url while configured_token_url keeps the + admin-entered value; the M2M spec must carry it so egress can mint.""" + spec = to_server_spec( + _server( + auth_type=MCPAuth.oauth2, + oauth2_flow="client_credentials", + url="https://up.example.com/mcp", + token_url=None, + configured_token_url="https://idp.example.com/token", + client_id="cid", + client_secret="csec", + ) + ) + assert spec is not None + assert isinstance(spec.config, ClientCredentialsConfig) + assert spec.config.token_url == "https://idp.example.com/token" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_authz_code_refresher.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_authz_code_refresher.py index ab414d1e8a4..bb2f2ff8b02 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_authz_code_refresher.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_authz_code_refresher.py @@ -20,8 +20,10 @@ class _Server: upstream_resource=None, url=None, server_id="srv", + configured_token_url=None, ): self.token_url = token_url + self.configured_token_url = configured_token_url self.client_id = client_id self.client_secret = client_secret self.token_endpoint_auth_method = token_endpoint_auth_method @@ -29,6 +31,10 @@ class _Server: self.url = url self.server_id = server_id + @property + def effective_token_url(self): + return self.token_url or self.configured_token_url + def _lookup(server): return lambda server_id: server @@ -262,3 +268,22 @@ async def test_returned_scope_overrides_prior_when_present(): assert token is not None assert token.scopes == ("read",) # a present scope replaces the prior grant assert persisted[0][5] == ("read",) + + +@pytest.mark.asyncio +async def test_refresh_uses_admin_entered_token_url_when_issuer_yield_empties_resolved(): + """A pinned issuer empties the resolved token_url while configured_token_url keeps the + admin-entered value; the refresh grant must POST there instead of silently failing.""" + posted = [] + refresher = _refresher( + server=_Server(token_url=None, configured_token_url="https://idp.example.com/token"), + body={"access_token": "new-at", "expires_in": 3600}, + post_sink=posted, + ) + token = await refresher.refresh( + "alice", "srv", OAuthToken(access_token="old", refresh_token="old-rt") + ) + + assert token is not None + assert token.access_token == "new-at" + assert posted[0][0] == "https://idp.example.com/token" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_db_credentials.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_db_credentials.py index 50248e95ffa..4d9142ad4c5 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_db_credentials.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_db_credentials.py @@ -1337,3 +1337,28 @@ def test_mcp_oauth_token_identity_changes_when_only_upstream_resource_is_edited( assert mcp_oauth_token_identity(set_to_explicit) == mcp_oauth_token_identity( _identity_server(credentials={**creds, "upstream_resource": "api://audience-one"}) ) + + +@pytest.mark.asyncio +async def test_refresh_user_oauth_token_uses_admin_entered_token_url_when_issuer_yield_empties_resolved(monkeypatch): + """A pinned issuer empties the resolved token_url while configured_token_url keeps the + admin-entered value; the silent per-user refresh must POST there instead of bailing.""" + from litellm.proxy._types import MCPTransport + from litellm.types.mcp import MCPAuth + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + server = MCPServer( + server_id="srv-1", + name="test", + url="https://up.example.com/mcp", + transport=MCPTransport.http, + auth_type=MCPAuth.oauth2, + client_id="cid", + client_secret="csec", + token_url=None, + configured_token_url="https://idp.example.com/token", + ) + result, captured = await _run_refresh(monkeypatch, server) + + assert result is not None + assert captured["url"] == "https://idp.example.com/token" diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py index 72589fd8b3e..b1aa16a30c0 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_oauth2_token_cache.py @@ -392,3 +392,22 @@ async def test_invalidate_clears_every_identity_for_a_server(): assert refetched == "tok-after-invalidate" assert mock_client.post.call_count == 3 + + +@pytest.mark.asyncio +async def test_m2m_mint_uses_admin_entered_token_url_when_issuer_yield_empties_resolved(): + """A pinned issuer empties the resolved token_url while configured_token_url keeps the + admin-entered value; the client_credentials mint must POST there instead of raising.""" + server = _server(token_url=None, configured_token_url="https://auth.example.com/token") + cache = MCPOAuth2TokenCache() + mock_client = AsyncMock() + mock_client.post.return_value = _token_response("m2m-token-configured") + + with patch( + "litellm.proxy._experimental.mcp_server.oauth2_token_cache.get_async_httpx_client", + return_value=mock_client, + ): + result = await cache.async_get_token(server) + + assert result == "m2m-token-configured" + assert mock_client.post.call_args[0][0] == "https://auth.example.com/token" From 93e7e8d98012ff4a217b743c803989fe91b8c2a5 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 16:26:44 -0700 Subject: [PATCH 6/6] fix(mcp): token exchange rejoins discovery for a clientless DCR bridge missing its registration endpoint --- .../mcp_server/discoverable_endpoints.py | 14 ++++- .../mcp_server/test_discoverable_endpoints.py | 59 +++++++++++++++++++ 2 files changed, 72 insertions(+), 1 deletion(-) diff --git a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py index 46feeab4dc3..93b85edd88d 100644 --- a/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py +++ b/litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py @@ -967,7 +967,7 @@ async def exchange_token_with_server( if grant_type not in ("authorization_code", "refresh_token"): raise HTTPException(status_code=400, detail="Unsupported grant_type") - resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, lambda s: s.effective_token_url) + resolved_server: Final = await _server_with_oauth_endpoints(mcp_server, _token_flow_needed_endpoint) token_url: Final = resolved_server.effective_token_url if token_url is None: raise HTTPException( @@ -1664,6 +1664,18 @@ def _register_flow_needed_endpoint(mcp_server: MCPServer) -> str | None: return mcp_server.effective_authorization_url +def _token_flow_needed_endpoint(mcp_server: MCPServer) -> str | None: + """The token exchange's deferred-discovery join gate. The exchange's relay-vs-callback arm + (:func:`_dcr_bridge_relays_client_registration`) reads the registration url, so a clientless + DCR bridge rebuilt without its discovered registration endpoint must keep joining discovery + even when the token url already resolves; skipping it would select the gateway-callback arm + and the upstream would reject the code over a redirect_uri mismatch. Every other shape only + needs the token url.""" + if mcp_server.is_dcr_bridge and not mcp_server.client_id and mcp_server.effective_registration_url is None: + return None + return mcp_server.effective_token_url + + async def register_client_with_server( request: Request, mcp_server: MCPServer, diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py index b1a99b498e5..04fb17ee6aa 100644 --- a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py +++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py @@ -247,6 +247,65 @@ async def test_register_route_bridge_missing_registration_url_joins_discovery(): assert json.loads(response.body.decode("utf-8"))["client_id"] == "generated-client" +@pytest.mark.asyncio +async def test_token_route_bridge_missing_registration_url_joins_discovery(): + """A clientless DCR bridge rebuilt with an admin-entered token url but without its discovered + registration endpoint must rejoin discovery at the exchange: the relay-vs-callback arm hinges + on the registration url, so skipping discovery would swap the client's own redirect_uri for + the gateway callback and the upstream would reject the code.""" + from litellm.proxy._experimental.mcp_server import discoverable_endpoints + from litellm.proxy._experimental.mcp_server.mcp_server_manager import global_mcp_server_manager + from litellm.proxy._types import MCPTransport + from litellm.types.mcp_server.mcp_server_manager import MCPServer + + server = MCPServer( + server_id="bridge-partial-token-metadata", + name="bridge_partial_token_metadata", + server_name="bridge_partial_token_metadata", + alias="bridge_partial_token_metadata", + url="https://mcp.example.com/mcp", + transport=MCPTransport.http, + auth_type=MCPAuth.true_passthrough, + dcr_bridge=True, + client_id=None, + authorization_url="https://idp.example.com/authorize", + token_url="https://idp.example.com/token", + ) + global_mcp_server_manager.registry[server.server_id] = server + global_mcp_server_manager._set_oauth_discovery_deferred(server.server_id, True) + request = _mock_callback_request("https://litellm.example.com/") + fake_http_response = MagicMock() + fake_http_response.json.return_value = {"access_token": "tok", "token_type": "Bearer"} + fake_http_response.raise_for_status = MagicMock() + fake_http_client = MagicMock() + fake_http_client.post = AsyncMock(return_value=fake_http_response) + + with ( + patch.object( # test-quality-ok: innermost discovery seam on a module-global manager; the route-to-exchange join under test stays real + global_mcp_server_manager, + "_discover_oauth_metadata_for_server", + new=AsyncMock(return_value=_resolved_oauth_metadata()), + ) as discovery, + patch.object( # test-quality-ok: keeps the upstream token POST off the network so its redirect_uri arm can be asserted + discoverable_endpoints, + "get_async_httpx_client", + new=lambda llm_provider: fake_http_client, + ), + ): + response = await discoverable_endpoints.token_endpoint( + request=request, + grant_type="authorization_code", + code="upstream-code", + redirect_uri="https://client.example.com/cb", + client_id="dcr-client-id", + mcp_server_name=server.server_name, + ) + + discovery.assert_awaited_once_with(server) + assert response.status_code == 200 + assert fake_http_client.post.await_args.kwargs["data"]["redirect_uri"] == "https://client.example.com/cb" + + @pytest.fixture def trust_xff(): """Force ``IPAddressUtils.is_request_from_trusted_proxy`` to True.