Upload mcp_server.py

The tool surface goes from 12 to 6: search, search_arxiv, search_google_scholar, search_patents, scrap_patent, scrap_patents. The exclusions live in one EXCLUDED_ROUTES list at the top of the file with the reasoning next to each pattern, so it's easy to adjust. MCP_EXPOSE_ALL=1 puts all 12 back without a code change — I verified both paths (6 tools by default, 12 with the flag set).

I also rewrote the server instructions, since the old ones described ops_* tools the model can no longer see. The new text tells the agent that search and search_patents already exhaust every backend internally, so a "not found" is genuine and it should move to the next id rather than hunting for a lower-level tool — which is exactly the loop that produced your JP7904662B2 traceback.

All the /ops/* REST endpoints still work exactly as before; this only changes what an LLM is offered.

One caveat worth knowing: after you redeploy, Claude may keep serving a cached tool list for a few minutes. If you still see the old 12 tools, toggle the connector off and on.
Claude

mcp_server.py

@@ -24,6 +24,8 @@ Configuration (all optional, read from the environment):
                        is what you want behind the HF Spaces proxy.
     MCP_ALLOWED_HOSTS  Comma-separated Host allow-list. Default "*".
     MCP_ALLOWED_ORIGINS Comma-separated Origin allow-list. Default "*".
+    MCP_EXPOSE_ALL     "1" to expose every endpoint as a tool instead of the
+                       curated set (see EXCLUDED_ROUTES below).
 """
 
 from __future__ import annotations
@@ -37,12 +39,33 @@ from fastapi.routing import APIRoute
 from fastapi.utils import generate_unique_id
 
 from fastmcp import FastMCP
+from fastmcp.server.providers.openapi import MCPType, RouteMap
 from fastmcp.utilities.lifespan import combine_lifespans
 
 logger = logging.getLogger(__name__)
 
 MCP_SERVER_NAME = "SERPent"
 
+# Endpoints kept out of the MCP tool surface. They stay fully available over
+# REST — this only controls what an LLM sees.
+#
+# Fewer, well-differentiated tools measurably improve tool selection, and both
+# groups below are redundant *for an agent*:
+#
+#   /ops/*        The regular `scrap_patent` / `search_patents` tools already
+#                 fall back to EPO OPS on their own. Calling the OPS tools
+#                 directly skips that fallback, so a patent missing from OPS
+#                 becomes a hard error instead of a Google Patents hit.
+#   search_brave  These are `search` with the fallback chain removed. `search`
+#   search_bing   already tries DuckDuckGo, then Brave, then Bing, so exposing
+#   search_duck   them separately only invites the model to pick a worse path.
+#
+# Set MCP_EXPOSE_ALL=1 to expose everything again, or edit this list.
+EXCLUDED_ROUTES = [
+    r"^/ops/.*",
+    r"^/serp/search_(brave|bing|duck)$",
+]
+
 MCP_INSTRUCTIONS = """\
 SERPent gives you web, academic and patent search plus full-text patent retrieval.
 
@@ -59,9 +82,11 @@ Every search tool accepts a LIST of queries and runs them concurrently, so send
 all your query variations in one call rather than making several calls. Set
 `n_results` to control how many hits come back per query.
 
-The `ops_*` tools hit the official EPO OPS API directly and only work when the
-deployment has OPS credentials configured; the regular tools already fall back
-to OPS on their own, so prefer them unless you specifically need EPO data.
+`search` and `search_patents` already fall back across several backends
+internally, including the official EPO OPS patent API, so a single call is the
+most thorough option available — there is no lower-level tool to reach for when
+one comes back empty. A patent that returns "not found" is genuinely absent
+from every backend; move on to the next id rather than retrying.
 """
 
 
@@ -117,11 +142,21 @@ def _tool_names(app: FastAPI) -> dict[str, str]:
 
 
 def build_mcp_server(app: FastAPI) -> FastMCP:
-    """Build an MCP server exposing every route of `app` as a tool."""
+    """Build an MCP server exposing the app's routes as tools."""
+    if _env_flag("MCP_EXPOSE_ALL", False):
+        route_maps = None
+        logger.info("MCP exposing all endpoints (MCP_EXPOSE_ALL=1).")
+    else:
+        route_maps = [
+            RouteMap(pattern=pattern, mcp_type=MCPType.EXCLUDE)
+            for pattern in EXCLUDED_ROUTES
+        ]
+
     mcp = FastMCP.from_fastapi(
         app=app,
         name=MCP_SERVER_NAME,
         mcp_names=_tool_names(app),
+        route_maps=route_maps,
         # Requests are dispatched straight back into this same app in-process.
         httpx_client_kwargs={"timeout": 300.0},
     )