feat(mcp): backport the OBO token endpoint discovery stack (#31526, #31622, #31762) to 1.91.0rc1

litellm/experimental_mcp_client/client.py

@@ -520,13 +520,28 @@ async def _list_tools_operation(session: ClientSession):
             # Return empty list instead of raising to allow graceful degradation
             return []
 
+    @staticmethod
+    def error_tool_result(exc: Exception) -> MCPCallToolResult:
+        """The error result ``call_tool`` returns when it swallows a failure (no re-execution)."""
+        return MCPCallToolResult(
+            content=[TextContent(type="text", text=f"{type(exc).__name__}: {str(exc)}")],
+            isError=True,
+        )
+
     async def call_tool(
         self,
         call_tool_request_params: MCPCallToolRequestParams,
         host_progress_callback: Optional[Callable] = None,
+        raise_on_error: bool = False,
     ) -> MCPCallToolResult:
         """
         Call an MCP Tool.
+
+        Args:
+            raise_on_error: When True, re-raise the underlying exception instead of returning an
+                ``isError=True`` result. The token-exchange (OBO) tool-call path uses this to detect
+                an upstream 401 so it can re-mint the exchanged token and retry once; every other
+                caller keeps the default and gets graceful ``isError`` degradation.
         """
         verbose_logger.info(f"MCP client calling tool '{call_tool_request_params.name}'")
 
@@ -579,11 +594,10 @@ async def _call_tool_operation(session: ClientSession):
                     "MCP client detected broken connection/stream - "
                     "the MCP server may have crashed, disconnected, or timed out."
                 )
+            if raise_on_error:
+                raise
             # Return a default error result instead of raising
-            return MCPCallToolResult(
-                content=[TextContent(type="text", text=f"{error_type}: {str(e)}")],  # Empty content for error case
-                isError=True,
-            )
+            return self.error_tool_result(e)
 
     async def list_prompts(self) -> List[Prompt]:
         """List available prompts from the server."""

litellm/proxy/_experimental/mcp_server/discoverable_endpoints.py

@@ -1184,6 +1184,10 @@ async def _build_oauth_protected_resource_response(
             detail=(f"Upstream oauth-protected-resource metadata unavailable for MCP server {mcp_server.name!r}"),
         )
 
+    obo_response = _obo_protected_resource_response(mcp_server, resource_url)
+    if obo_response is not None:
+        return obo_response
+
     return {
         "authorization_servers": [
             (f"{request_base_url}/{mcp_server_name}" if mcp_server_name else f"{request_base_url}")
@@ -1193,6 +1197,51 @@ async def _build_oauth_protected_resource_response(
     }
 
 
+def _obo_protected_resource_response(mcp_server: Optional[MCPServer], resource_url: str) -> Optional[dict]:
+    """The OBO (token_exchange) PRM, or None when this server is not OBO / no issuer is configured.
+
+    The client SSOs with the IdP to obtain a subject token, which LiteLLM then exchanges, so discovery
+    points at the JWT-auth issuer(s) LiteLLM trusts (the same IdP that issues and validates the
+    subject), not the gateway. None falls the caller back to the gateway default so discovery still
+    returns metadata; it just can't name the IdP.
+    """
+    if mcp_server is None or mcp_server.auth_type != MCPAuth.oauth2_token_exchange:
+        return None
+    issuers = _jwt_auth_issuers()
+    if not issuers:
+        return None
+    return {
+        "authorization_servers": issuers,
+        "resource": resource_url,
+        "scopes_supported": (mcp_server.scopes if mcp_server.scopes else []),
+    }
+
+
+def _jwt_auth_issuers() -> list:
+    """The OAuth issuer identifier(s) LiteLLM's JWT auth trusts, for the OBO PRM authorization_servers.
+
+    In token_exchange the IdP that issues the subject JWT is the same one LiteLLM validates it
+    against, so OBO discovery points clients at the JWT-auth issuer to obtain a subject token.
+    Sourced from ``JWT_ISSUER`` and any configured ``litellm_jwtauth.issuers``.
+    """
+    import os  # noqa: PLC0415
+
+    from litellm.proxy.proxy_server import general_settings  # noqa: PLC0415
+
+    issuers: list = []
+    env_issuer = os.getenv("JWT_ISSUER")
+    if env_issuer:
+        issuers.append(env_issuer)
+
+    jwtauth = general_settings.get("litellm_jwtauth") if isinstance(general_settings, dict) else None
+    raw_issuers = jwtauth.get("issuers") if isinstance(jwtauth, dict) else getattr(jwtauth, "issuers", None)
+    for cfg in raw_issuers or []:
+        issuer = cfg.get("issuer") if isinstance(cfg, dict) else getattr(cfg, "issuer", None)
+        if issuer and issuer not in issuers:
+            issuers.append(issuer)
+    return issuers
+
+
 # Standard MCP pattern: /.well-known/oauth-protected-resource/mcp/{server_name}
 # This is the pattern expected by standard MCP clients (mcp-inspector, VSCode Copilot)
 @router.get(