test(e2e): cover MCP api_key tools/list + tools/call over a virtual key

Adds a live e2e for the MCP gateway's api_key auth family: a LiteLLM virtual
key discovers a config-registered upstream server's tools and invokes one, with
the gateway injecting the upstream credential. The upstream is incidental (the
hugging_face server today); the test runs hub_repo_search and asserts a Hub
search for GPT-2 under openai-community surfaces openai-community/gpt2, so a
broken grant, a dropped upstream credential, or a tool that never ran fails
rather than passing on a bare 200.

The server is owned by gateway config: the test looks it up by name, hard-fails
(never skips) when absent, and tears down only the key. Adds server_name to
McpServerRow so the lookup works.

Covers mcp.list_tools.api_key.succeeds and mcp.call_tool.api_key.succeeds.

tests/e2e/mcp/mcp_client.py

@@ -44,6 +44,7 @@ class McpServerNewResponse(BaseModel):
 
 class McpServerRow(BaseModel):
     server_id: str
+    server_name: str | None = None
     alias: str | None = None
     url: str | None = None
 

tests/e2e/mcp/test_mcp_api_key_tools_e2e.py

@@ -0,0 +1,87 @@
+"""Live e2e: the MCP gateway brokers a config-registered upstream server over a virtual key.
+
+Exercises the api_key auth family end to end: a LiteLLM virtual key (x-litellm-api-key)
+discovers an upstream MCP server's tools (tools/list) and invokes one (tools/call), with
+the gateway injecting the upstream credential itself. The upstream here is the
+config-registered "hugging_face" server; it is only the vehicle. The call runs
+hub_repo_search and asserts a Hub search for GPT-2 under the openai-community namespace
+surfaces openai-community/gpt2, so a broken grant, a dropped upstream credential, or a
+tool that never ran fails rather than passing on a bare 200. The server is owned by
+gateway config, so the test looks it up by name, hard-fails (never skips) when absent,
+and tears down only the virtual key.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from e2e_config import unique_marker
+from e2e_http import unwrap
+from lifecycle import ResourceManager
+from mcp_client import McpClient
+
+pytestmark = pytest.mark.e2e
+
+UPSTREAM_SERVER_NAME = "hugging_face"
+SEARCH_TOOL = "hub_repo_search"
+SEARCH_QUERY = "gpt2"
+SEARCH_AUTHOR = "openai-community"
+CANONICAL_REPO_ID = "openai-community/gpt2"
+
+
+def _find_server(client: McpClient, server_name: str) -> str:
+    """Return the server_id of a config-registered MCP server, or hard-fail.
+
+    Live e2e never skips for a missing upstream: a gateway without the server is a red
+    run so the operator knows the suite cannot prove the product path.
+    """
+    for row in client.registered_servers():
+        if row.server_name == server_name:
+            return row.server_id
+    pytest.fail(
+        f"gateway has no {server_name!r} MCP server registered; add it to the proxy "
+        "config and restart"
+    )
+
+
+class TestMcpApiKeyToolAccess:
+    @pytest.mark.covers("mcp.list_tools.api_key.succeeds", "mcp.call_tool.api_key.succeeds")
+    def test_virtual_key_lists_and_calls_upstream_tool(
+        self,
+        client: McpClient,
+        resources: ResourceManager,
+    ) -> None:
+        server_id = _find_server(client, UPSTREAM_SERVER_NAME)
+
+        key = client.generate_key(
+            user_id=f"e2e-mcp-api-key-{unique_marker()}",
+            mcp_servers=[server_id],
+        )
+        resources.defer(lambda: client.proxy.delete_key(key))
+
+        tools = unwrap(client.list_tools(key))
+        tool_name = tools.tool_name_containing(server_id, SEARCH_TOOL)
+        assert tool_name is not None, (
+            f"granted key never saw {SEARCH_TOOL} on server {UPSTREAM_SERVER_NAME!r}; "
+            f"tools={tools.tool_names_for_server(server_id)}"
+        )
+
+        call = unwrap(
+            client.call_tool(
+                key,
+                server_id=server_id,
+                name=tool_name,
+                arguments={
+                    "query": SEARCH_QUERY,
+                    "author": SEARCH_AUTHOR,
+                    "repo_types": ["model"],
+                    "limit": 10,
+                },
+            )
+        )
+        assert call.is_error is not True, f"{SEARCH_TOOL} errored: {call}"
+        body = call.all_text
+        assert CANONICAL_REPO_ID in body, (
+            f"{SEARCH_TOOL} for {SEARCH_QUERY!r} under {SEARCH_AUTHOR!r} must surface "
+            f"{CANONICAL_REPO_ID!r}; got: {body[:800]!r}"
+        )