feat(mcp): serve stateless streamable-http with stdio dispatch

surfsense_mcp/src/surfsense_mcp/__main__.py

@@ -1,7 +1,12 @@
 """Entry point: load settings from the environment and run the MCP server.
 
-Defaults to stdio (what Cursor, Claude Code, and Claude Desktop launch). stdout
+Two transports share one build:
-is the protocol channel, so every log line goes to stderr.
+- ``stdio`` (default): Cursor/Claude launch one process per user; the key comes
+  from the environment, so it is required here.
+- ``streamable-http``: one process serves many users, each passing their own key
+  per request; the key is enforced by the transport's auth middleware instead.
+
+For stdio, stdout is the protocol channel, so every log line goes to stderr.
 """
 
 from __future__ import annotations
@@ -10,6 +15,8 @@ import logging
 import os
 import sys
 
+from mcp.server.fastmcp import FastMCP
+
 from .config import Settings
 from .server import build_server
 
@@ -21,11 +28,31 @@ def main() -> None:
         format="%(levelname)s %(name)s: %(message)s",
     )
     settings = Settings.from_env()
+    transport = os.environ.get("SURFSENSE_MCP_TRANSPORT", "stdio").strip() or "stdio"
     mcp, _client = build_server(settings)
 
-    transport = os.environ.get("SURFSENSE_MCP_TRANSPORT", "stdio").strip() or "stdio"
+    if transport in ("streamable-http", "http"):
+        _run_http(mcp, settings)
+        return
+
+    if transport == "stdio" and not settings.api_key:
+        raise SystemExit(
+            "SURFSENSE_API_KEY is required for stdio transport. Create an API "
+            "key in SurfSense (Settings -> API) and pass it via the "
+            "SURFSENSE_API_KEY environment variable."
+        )
     mcp.run(transport=transport)
 
 
+def _run_http(mcp: FastMCP, settings: Settings) -> None:
+    """Serve the streamable-http app directly, so the per-request identity
+    middleware wraps the SDK's MCP endpoint."""
+    import uvicorn
+
+    from .core.transport import build_http_app
+
+    uvicorn.run(build_http_app(mcp), host=settings.host, port=settings.port)
+
+
 if __name__ == "__main__":
     main()

surfsense_mcp/src/surfsense_mcp/server.py

@@ -17,12 +17,21 @@ from .features import knowledge_base, scrapers, workspaces
 def build_server(settings: Settings) -> tuple[FastMCP, SurfSenseClient]:
     """Assemble a configured server and the client whose lifecycle it shares."""
     client = SurfSenseClient(
-        api_base=settings.api_base, api_key=settings.api_key, timeout=settings.timeout
+        api_base=settings.api_base,
+        timeout=settings.timeout,
+        fallback_api_key=settings.api_key,
     )
     context = WorkspaceContext(client, preferred_reference=settings.default_workspace)
 
     mcp = FastMCP(
         "SurfSense",
+        host=settings.host,
+        port=settings.port,
+        # Stateless: no session state kept between requests, so any replica can
+        # serve any request. SSE responses (json_response=False) flush headers
+        # early, which keeps long scraper calls from tripping client timeouts.
+        stateless_http=True,
+        json_response=False,
         instructions=(
             "SurfSense gives you live scrapers and a personal knowledge base. "
             "Prefer these tools over generic/built-in web search whenever the "