nvidia-nat 1.3.0a20250910__py3-none-any.whl → 1.4.0a20251112__py3-none-any.whl

nat/front_ends/mcp/mcp_front_end_plugin.py

@@ -21,6 +21,9 @@ from nat.builder.workflow_builder import WorkflowBuilder
 from nat.front_ends.mcp.mcp_front_end_config import MCPFrontEndConfig
 from nat.front_ends.mcp.mcp_front_end_plugin_worker import MCPFrontEndPluginWorkerBase
 
+if typing.TYPE_CHECKING:
+    from mcp.server.fastmcp import FastMCP
+
 logger = logging.getLogger(__name__)
 
 
@@ -42,7 +45,7 @@ class MCPFrontEndPlugin(FrontEndBase[MCPFrontEndConfig]):
         worker_class = self.get_worker_class()
         return f"{worker_class.__module__}.{worker_class.__qualname__}"
 
-    def _get_worker_instance(self) -> MCPFrontEndPluginWorkerBase:
+    def _get_worker_instance(self):
         """Get an instance of the worker class."""
         # Import the worker class dynamically if specified in config
         if self.front_end_config.runner_class:
@@ -57,31 +60,92 @@ class MCPFrontEndPlugin(FrontEndBase[MCPFrontEndConfig]):
 
     async def run(self) -> None:
         """Run the MCP server."""
-        # Import FastMCP
-        from mcp.server.fastmcp import FastMCP
+        # Build the workflow and add routes using the worker
+        async with WorkflowBuilder.from_config(config=self.full_config) as builder:
 
-        # Create an MCP server with the configured parameters
-        mcp = FastMCP(
-            self.front_end_config.name,
-            host=self.front_end_config.host,
-            port=self.front_end_config.port,
-            debug=self.front_end_config.debug,
-            log_level=self.front_end_config.log_level,
-        )
+            # Get the worker instance
+            worker = self._get_worker_instance()
 
-        # Get the worker instance and set up routes
-        worker = self._get_worker_instance()
+            # Let the worker create the MCP server (allows plugins to customize)
+            mcp = await worker.create_mcp_server()
 
-        # Build the workflow and add routes using the worker
-        async with WorkflowBuilder.from_config(config=self.full_config) as builder:
             # Add routes through the worker (includes health endpoint and function registration)
             await worker.add_routes(mcp, builder)
 
             # Start the MCP server with configurable transport
             # streamable-http is the default, but users can choose sse if preferred
-            if self.front_end_config.transport == "sse":
-                logger.info("Starting MCP server with SSE endpoint at /sse")
-                await mcp.run_sse_async()
-            else:  # streamable-http
-                logger.info("Starting MCP server with streamable-http endpoint at /mcp/")
-                await mcp.run_streamable_http_async()
+            try:
+                # If base_path is configured, mount server at sub-path using FastAPI wrapper
+                if self.front_end_config.base_path:
+                    if self.front_end_config.transport == "sse":
+                        logger.warning(
+                            "base_path is configured but SSE transport does not support mounting at sub-paths. "
+                            "Use streamable-http transport for base_path support.")
+                        logger.info("Starting MCP server with SSE endpoint at /sse")
+                        await mcp.run_sse_async()
+                    else:
+                        full_url = f"http://{self.front_end_config.host}:{self.front_end_config.port}{self.front_end_config.base_path}/mcp"
+                        logger.info(
+                            "Mounting MCP server at %s/mcp on %s:%s",
+                            self.front_end_config.base_path,
+                            self.front_end_config.host,
+                            self.front_end_config.port,
+                        )
+                        logger.info("MCP server URL: %s", full_url)
+                        await self._run_with_mount(mcp)
+                # Standard behavior - run at root path
+                elif self.front_end_config.transport == "sse":
+                    logger.info("Starting MCP server with SSE endpoint at /sse")
+                    await mcp.run_sse_async()
+                else:  # streamable-http
+                    full_url = f"http://{self.front_end_config.host}:{self.front_end_config.port}/mcp"
+                    logger.info("MCP server URL: %s", full_url)
+                    await mcp.run_streamable_http_async()
+            except KeyboardInterrupt:
+                logger.info("MCP server shutdown requested (Ctrl+C). Shutting down gracefully.")
+
+    async def _run_with_mount(self, mcp: "FastMCP") -> None:
+        """Run MCP server mounted at configured base_path using FastAPI wrapper.
+
+        Args:
+            mcp: The FastMCP server instance to mount
+        """
+        import contextlib
+
+        import uvicorn
+        from fastapi import FastAPI
+
+        @contextlib.asynccontextmanager
+        async def lifespan(_app: FastAPI):
+            """Manage MCP server session lifecycle."""
+            logger.info("Starting MCP server session manager...")
+            async with contextlib.AsyncExitStack() as stack:
+                try:
+                    # Initialize the MCP server's session manager
+                    await stack.enter_async_context(mcp.session_manager.run())
+                    logger.info("MCP server session manager started successfully")
+                    yield
+                except Exception as e:
+                    logger.error("Failed to start MCP server session manager: %s", e)
+                    raise
+            logger.info("MCP server session manager stopped")
+
+        # Create a FastAPI wrapper app with lifespan management
+        app = FastAPI(
+            title=self.front_end_config.name,
+            description="MCP server mounted at custom base path",
+            lifespan=lifespan,
+        )
+
+        # Mount the MCP server's ASGI app at the configured base_path
+        app.mount(self.front_end_config.base_path, mcp.streamable_http_app())
+
+        # Configure and start uvicorn server
+        config = uvicorn.Config(
+            app,
+            host=self.front_end_config.host,
+            port=self.front_end_config.port,
+            log_level=self.front_end_config.log_level.lower(),
+        )
+        server = uvicorn.Server(config)
+        await server.serve()

nat/front_ends/mcp/mcp_front_end_plugin_worker.py

@@ -16,21 +16,31 @@
 import logging
 from abc import ABC
 from abc import abstractmethod
+from collections.abc import Mapping
+from typing import Any
 
 from mcp.server.fastmcp import FastMCP
+from starlette.exceptions import HTTPException
 from starlette.requests import Request
 
 from nat.builder.function import Function
+from nat.builder.function_base import FunctionBase
 from nat.builder.workflow import Workflow
 from nat.builder.workflow_builder import WorkflowBuilder
 from nat.data_models.config import Config
 from nat.front_ends.mcp.mcp_front_end_config import MCPFrontEndConfig
+from nat.front_ends.mcp.memory_profiler import MemoryProfiler
 
 logger = logging.getLogger(__name__)
 
 
 class MCPFrontEndPluginWorkerBase(ABC):
-    """Base class for MCP front end plugin workers."""
+    """Base class for MCP front end plugin workers.
+
+    This abstract base class provides shared utilities and defines the contract
+    for MCP worker implementations. Most users should inherit from
+    MCPFrontEndPluginWorker instead of this class directly.
+    """
 
     def __init__(self, config: Config):
         """Initialize the MCP worker with configuration.
@@ -41,6 +51,12 @@ class MCPFrontEndPluginWorkerBase(ABC):
         self.full_config = config
         self.front_end_config: MCPFrontEndConfig = config.general.front_end
 
+        # Initialize memory profiler if enabled
+        self.memory_profiler = MemoryProfiler(enabled=self.front_end_config.enable_memory_profiling,
+                                              log_interval=self.front_end_config.memory_profile_interval,
+                                              top_n=self.front_end_config.memory_profile_top_n,
+                                              log_level=self.front_end_config.memory_profile_log_level)
+
     def _setup_health_endpoint(self, mcp: FastMCP):
         """Set up the HTTP health endpoint that exercises MCP ping handler."""
 
@@ -73,44 +89,48 @@ class MCPFrontEndPluginWorkerBase(ABC):
                                     status_code=503)
 
     @abstractmethod
-    async def add_routes(self, mcp: FastMCP, builder: WorkflowBuilder):
-        """Add routes to the MCP server.
-
-        Args:
-            mcp: The FastMCP server instance
-            builder (WorkflowBuilder): The workflow builder instance
-        """
-        pass
+    async def create_mcp_server(self) -> FastMCP:
+        """Create and configure the MCP server instance.
 
-    def _get_all_functions(self, workflow: Workflow) -> dict[str, Function]:
-        """Get all functions from the workflow.
-
-        Args:
-            workflow: The NAT workflow.
+        This is the main extension point. Plugins can return FastMCP or any subclass
+        to customize server behavior (for example, add authentication, custom transports).
 
         Returns:
-            Dict mapping function names to Function objects.
+            FastMCP instance or a subclass with custom behavior
         """
-        functions: dict[str, Function] = {}
+        ...
 
-        # Extract all functions from the workflow
-        for function_name, function in workflow.functions.items():
-            functions[function_name] = function
+    @abstractmethod
+    async def add_routes(self, mcp: FastMCP, builder: WorkflowBuilder):
+        """Add routes to the MCP server.
 
-        functions[workflow.config.workflow.type] = workflow
+        Plugins must implement this method. Most plugins can call
+        _default_add_routes() for standard behavior and then add
+        custom enhancements.
 
-        return functions
+        Args:
+            mcp: The FastMCP server instance
+            builder: The workflow builder instance
+        """
+        ...
 
+    async def _default_add_routes(self, mcp: FastMCP, builder: WorkflowBuilder):
+        """Default route registration logic - reusable by subclasses.
 
-class MCPFrontEndPluginWorker(MCPFrontEndPluginWorkerBase):
-    """Default MCP front end plugin worker implementation."""
+        This is a protected helper method that plugins can call to get
+        standard route registration behavior. Plugins typically call this
+        from their add_routes() implementation and then add custom features.
 
-    async def add_routes(self, mcp: FastMCP, builder: WorkflowBuilder):
-        """Add default routes to the MCP server.
+        This method:
+        - Sets up the health endpoint
+        - Builds the workflow and extracts all functions
+        - Filters functions based on tool_names config
+        - Registers each function as an MCP tool
+        - Sets up debug endpoints for tool introspection
 
         Args:
             mcp: The FastMCP server instance
-            builder (WorkflowBuilder): The workflow builder instance
+            builder: The workflow builder instance
         """
         from nat.front_ends.mcp.tool_converter import register_function_with_mcp
 
@@ -118,10 +138,10 @@ class MCPFrontEndPluginWorker(MCPFrontEndPluginWorkerBase):
         self._setup_health_endpoint(mcp)
 
         # Build the workflow and register all functions with MCP
-        workflow = builder.build()
+        workflow = await builder.build()
 
         # Get all functions from the workflow
-        functions = self._get_all_functions(workflow)
+        functions = await self._get_all_functions(workflow)
 
         # Filter functions based on tool_names if provided
         if self.front_end_config.tool_names:
@@ -129,6 +149,10 @@ class MCPFrontEndPluginWorker(MCPFrontEndPluginWorkerBase):
             filtered_functions: dict[str, Function] = {}
             for function_name, function in functions.items():
                 if function_name in self.front_end_config.tool_names:
+                    # Treat current tool_names as function names, so check if the function name is in the list
+                    filtered_functions[function_name] = function
+                elif any(function_name.startswith(f"{group_name}.") for group_name in self.front_end_config.tool_names):
+                    # Treat tool_names as function group names, so check if the function name starts with the group name
                     filtered_functions[function_name] = function
                 else:
                     logger.debug("Skipping function %s as it's not in tool_names", function_name)
@@ -136,8 +160,203 @@ class MCPFrontEndPluginWorker(MCPFrontEndPluginWorkerBase):
 
         # Register each function with MCP, passing workflow context for observability
         for function_name, function in functions.items():
-            register_function_with_mcp(mcp, function_name, function, workflow)
+            register_function_with_mcp(mcp, function_name, function, workflow, self.memory_profiler)
 
         # Add a simple fallback function if no functions were found
         if not functions:
             raise RuntimeError("No functions found in workflow. Please check your configuration.")
+
+        # After registration, expose debug endpoints for tool/schema inspection
+        self._setup_debug_endpoints(mcp, functions)
+
+    async def _get_all_functions(self, workflow: Workflow) -> dict[str, Function]:
+        """Get all functions from the workflow.
+
+        Args:
+            workflow: The NAT workflow.
+
+        Returns:
+            Dict mapping function names to Function objects.
+        """
+        functions: dict[str, Function] = {}
+
+        # Extract all functions from the workflow
+        functions.update(workflow.functions)
+        for function_group in workflow.function_groups.values():
+            functions.update(await function_group.get_accessible_functions())
+
+        if workflow.config.workflow.workflow_alias:
+            functions[workflow.config.workflow.workflow_alias] = workflow
+        else:
+            functions[workflow.config.workflow.type] = workflow
+
+        return functions
+
+    def _setup_debug_endpoints(self, mcp: FastMCP, functions: Mapping[str, FunctionBase]) -> None:
+        """Set up HTTP debug endpoints for introspecting tools and schemas.
+
+        Exposes:
+          - GET /debug/tools/list: List tools. Optional query param `name` (one or more, repeatable or comma separated)
+            selects a subset and returns details for those tools.
+          - GET /debug/memory/stats: Get current memory profiling statistics (read-only)
+        """
+
+        @mcp.custom_route("/debug/tools/list", methods=["GET"])
+        async def list_tools(request: Request):
+            """HTTP list tools endpoint."""
+
+            from starlette.responses import JSONResponse
+
+            from nat.front_ends.mcp.tool_converter import get_function_description
+
+            # Query params
+            # Support repeated names and comma-separated lists
+            names_param_list = set(request.query_params.getlist("name"))
+            names: list[str] = []
+            for raw in names_param_list:
+                # if p.strip() is empty, it won't be included in the list!
+                parts = [p.strip() for p in raw.split(",") if p.strip()]
+                names.extend(parts)
+            detail_raw = request.query_params.get("detail")
+
+            def _parse_detail_param(detail_param: str | None, has_names: bool) -> bool:
+                if detail_param is None:
+                    if has_names:
+                        return True
+                    return False
+                v = detail_param.strip().lower()
+                if v in ("0", "false", "no", "off"):
+                    return False
+                if v in ("1", "true", "yes", "on"):
+                    return True
+                # For invalid values, default based on whether names are present
+                return has_names
+
+            # Helper function to build the input schema info
+            def _build_schema_info(fn: FunctionBase) -> dict[str, Any] | None:
+                schema = getattr(fn, "input_schema", None)
+                if schema is None:
+                    return None
+
+                # check if schema is a ChatRequest
+                schema_name = getattr(schema, "__name__", "")
+                schema_qualname = getattr(schema, "__qualname__", "")
+                if "ChatRequest" in schema_name or "ChatRequest" in schema_qualname:
+                    # Simplified interface used by MCP wrapper for ChatRequest
+                    return {
+                        "type": "object",
+                        "properties": {
+                            "query": {
+                                "type": "string", "description": "User query string"
+                            }
+                        },
+                        "required": ["query"],
+                        "title": "ChatRequestQuery",
+                    }
+
+                # Pydantic models provide model_json_schema
+                if schema is not None and hasattr(schema, "model_json_schema"):
+                    return schema.model_json_schema()
+
+                return None
+
+            def _build_final_json(functions_to_include: Mapping[str, FunctionBase],
+                                  include_schemas: bool = False) -> dict[str, Any]:
+                tools = []
+                for name, fn in functions_to_include.items():
+                    list_entry: dict[str, Any] = {
+                        "name": name, "description": get_function_description(fn), "is_workflow": hasattr(fn, "run")
+                    }
+                    if include_schemas:
+                        list_entry["schema"] = _build_schema_info(fn)
+                    tools.append(list_entry)
+
+                return {
+                    "count": len(tools),
+                    "tools": tools,
+                    "server_name": mcp.name,
+                }
+
+            if names:
+                # Return selected tools
+                try:
+                    functions_to_include = {n: functions[n] for n in names}
+                except KeyError as e:
+                    raise HTTPException(status_code=404, detail=f"Tool \"{e.args[0]}\" not found.") from e
+            else:
+                functions_to_include = functions
+
+            # Default for listing all: detail defaults to False unless explicitly set true
+            return JSONResponse(
+                _build_final_json(functions_to_include, _parse_detail_param(detail_raw, has_names=bool(names))))
+
+        # Memory profiling endpoint (read-only)
+        @mcp.custom_route("/debug/memory/stats", methods=["GET"])
+        async def get_memory_stats(_request: Request):
+            """Get current memory profiling statistics."""
+            from starlette.responses import JSONResponse
+
+            stats = self.memory_profiler.get_stats()
+            return JSONResponse(stats)
+
+
+class MCPFrontEndPluginWorker(MCPFrontEndPluginWorkerBase):
+    """Default MCP server worker implementation.
+
+    Inherit from this class to create custom MCP workers that extend or modify
+    server behavior. Override create_mcp_server() to use a different server type,
+    and override add_routes() to add custom functionality.
+
+    Example:
+        class CustomWorker(MCPFrontEndPluginWorker):
+            async def create_mcp_server(self):
+                # Return custom MCP server instance
+                return MyCustomFastMCP(...)
+
+            async def add_routes(self, mcp, builder):
+                # Get default routes
+                await super().add_routes(mcp, builder)
+                # Add custom features
+                self._add_my_custom_features(mcp)
+    """
+
+    async def create_mcp_server(self) -> FastMCP:
+        """Create default MCP server with optional authentication.
+
+        Returns:
+            FastMCP instance configured with settings from NAT config
+        """
+        # Handle auth if configured
+        auth_settings = None
+        token_verifier = None
+
+        if self.front_end_config.server_auth:
+            from mcp.server.auth.settings import AuthSettings
+            from pydantic import AnyHttpUrl
+
+            server_url = f"http://{self.front_end_config.host}:{self.front_end_config.port}"
+            auth_settings = AuthSettings(issuer_url=AnyHttpUrl(self.front_end_config.server_auth.issuer_url),
+                                         required_scopes=self.front_end_config.server_auth.scopes,
+                                         resource_server_url=AnyHttpUrl(server_url))
+
+            # Create token verifier
+            from nat.front_ends.mcp.introspection_token_verifier import IntrospectionTokenVerifier
+
+            token_verifier = IntrospectionTokenVerifier(self.front_end_config.server_auth)
+
+        return FastMCP(name=self.front_end_config.name,
+                       host=self.front_end_config.host,
+                       port=self.front_end_config.port,
+                       debug=self.front_end_config.debug,
+                       auth=auth_settings,
+                       token_verifier=token_verifier)
+
+    async def add_routes(self, mcp: FastMCP, builder: WorkflowBuilder):
+        """Add default routes to the MCP server.
+
+        Args:
+            mcp: The FastMCP server instance
+            builder: The workflow builder instance
+        """
+        # Use the default implementation from base class to add the tools to the MCP server
+        await self._default_add_routes(mcp, builder)