nvidia-nat 1.2.0__py3-none-any.whl

nat/front_ends/fastapi/register.py

@@ -0,0 +1,25 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from nat.cli.register_workflow import register_front_end
+from nat.data_models.config import Config
+from nat.front_ends.fastapi.fastapi_front_end_config import FastApiFrontEndConfig
+
+
+@register_front_end(config_type=FastApiFrontEndConfig)
+async def register_fastapi_front_end(config: FastApiFrontEndConfig, full_config: Config):
+    from nat.front_ends.fastapi.fastapi_front_end_plugin import FastApiFrontEndPlugin
+
+    yield FastApiFrontEndPlugin(full_config=full_config)

nat/front_ends/fastapi/response_helpers.py

@@ -0,0 +1,195 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import asyncio
+import typing
+from collections.abc import AsyncGenerator
+
+from nat.data_models.api_server import ResponseIntermediateStep
+from nat.data_models.api_server import ResponsePayloadOutput
+from nat.data_models.api_server import ResponseSerializable
+from nat.data_models.step_adaptor import StepAdaptorConfig
+from nat.front_ends.fastapi.intermediate_steps_subscriber import pull_intermediate
+from nat.front_ends.fastapi.step_adaptor import StepAdaptor
+from nat.runtime.session import SessionManager
+from nat.utils.producer_consumer_queue import AsyncIOProducerConsumerQueue
+
+
+async def generate_streaming_response_as_str(payload: typing.Any,
+                                             *,
+                                             session_manager: SessionManager,
+                                             streaming: bool,
+                                             step_adaptor: StepAdaptor = StepAdaptor(StepAdaptorConfig()),
+                                             result_type: type | None = None,
+                                             output_type: type | None = None) -> AsyncGenerator[str]:
+
+    async for item in generate_streaming_response(payload,
+                                                  session_manager=session_manager,
+                                                  streaming=streaming,
+                                                  step_adaptor=step_adaptor,
+                                                  result_type=result_type,
+                                                  output_type=output_type):
+
+        if (isinstance(item, ResponseSerializable)):
+            yield item.get_stream_data()
+        else:
+            raise ValueError("Unexpected item type in stream. Expected ChatResponseSerializable, got: " +
+                             str(type(item)))
+
+
+async def generate_streaming_response(payload: typing.Any,
+                                      *,
+                                      session_manager: SessionManager,
+                                      streaming: bool,
+                                      step_adaptor: StepAdaptor = StepAdaptor(StepAdaptorConfig()),
+                                      result_type: type | None = None,
+                                      output_type: type | None = None) -> AsyncGenerator[ResponseSerializable]:
+
+    async with session_manager.run(payload) as runner:
+
+        q: AsyncIOProducerConsumerQueue[ResponseSerializable] = AsyncIOProducerConsumerQueue()
+
+        # Start the intermediate stream
+        intermediate_complete = await pull_intermediate(q, step_adaptor)
+
+        async def pull_result():
+            if session_manager.workflow.has_streaming_output and streaming:
+                async for chunk in runner.result_stream(to_type=output_type):
+                    await q.put(chunk)
+            else:
+                result = await runner.result(to_type=result_type)
+                await q.put(runner.convert(result, output_type))
+
+            # Wait until the intermediate subscription is done before closing q
+            # But we have no direct "intermediate_done" reference here
+            # because it's encapsulated in pull_intermediate. So we can do:
+            #    await some_event.wait()
+            # If needed. Alternatively, you can skip that if the intermediate
+            # subscriber won't block the main flow.
+            #
+            # For example, if you *need* to guarantee the subscriber is done before
+            # closing the queue, you can structure the code to store or return
+            # the 'intermediate_done' event from pull_intermediate.
+            #
+
+            await intermediate_complete.wait()
+
+            await q.close()
+
+        try:
+            # Start the result stream
+            asyncio.create_task(pull_result())
+
+            async for item in q:
+
+                if (isinstance(item, ResponseSerializable)):
+                    yield item
+                else:
+                    yield ResponsePayloadOutput(payload=item)
+        except Exception as e:
+            # Handle exceptions here
+            raise e
+        finally:
+            await q.close()
+
+
+async def generate_single_response(
+    payload: typing.Any,
+    session_manager: SessionManager,
+    result_type: type | None = None,
+) -> typing.Any:
+    if (not session_manager.workflow.has_single_output):
+        raise ValueError("Cannot get a single output value for streaming workflows")
+
+    async with session_manager.run(payload) as runner:
+        return await runner.result(to_type=result_type)
+
+
+async def generate_streaming_response_full(payload: typing.Any,
+                                           *,
+                                           session_manager: SessionManager,
+                                           streaming: bool,
+                                           result_type: type | None = None,
+                                           output_type: type | None = None,
+                                           filter_steps: str | None = None) -> AsyncGenerator[ResponseSerializable]:
+    """
+    Similar to generate_streaming_response but provides raw ResponseIntermediateStep objects
+    without any step adaptor translations.
+    """
+    # Parse filter_steps into a set of allowed types if provided
+    # Special case: if filter_steps is "none", suppress all steps
+    allowed_types = None
+    if filter_steps:
+        if filter_steps.lower() == "none":
+            allowed_types = set()  # Empty set means no steps allowed
+        else:
+            allowed_types = set(filter_steps.split(','))
+
+    async with session_manager.run(payload) as runner:
+        q: AsyncIOProducerConsumerQueue[ResponseSerializable] = AsyncIOProducerConsumerQueue()
+
+        # Start the intermediate stream without step adaptor
+        intermediate_complete = await pull_intermediate(q, None)
+
+        async def pull_result():
+            if session_manager.workflow.has_streaming_output and streaming:
+                async for chunk in runner.result_stream(to_type=output_type):
+                    await q.put(chunk)
+            else:
+                result = await runner.result(to_type=result_type)
+                await q.put(runner.convert(result, output_type))
+
+            await intermediate_complete.wait()
+            await q.close()
+
+        try:
+            # Start the result stream
+            asyncio.create_task(pull_result())
+
+            async for item in q:
+                if (isinstance(item, ResponseIntermediateStep)):
+                    # Filter intermediate steps if filter_steps is provided
+                    if allowed_types is None or item.type in allowed_types:
+                        yield item
+                else:
+                    yield ResponsePayloadOutput(payload=item)
+        except Exception as e:
+            # Handle exceptions here
+            raise e
+        finally:
+            await q.close()
+
+
+async def generate_streaming_response_full_as_str(payload: typing.Any,
+                                                  *,
+                                                  session_manager: SessionManager,
+                                                  streaming: bool,
+                                                  result_type: type | None = None,
+                                                  output_type: type | None = None,
+                                                  filter_steps: str | None = None) -> AsyncGenerator[str]:
+    """
+    Similar to generate_streaming_response but converts the response to a string format.
+    """
+    async for item in generate_streaming_response_full(payload,
+                                                       session_manager=session_manager,
+                                                       streaming=streaming,
+                                                       result_type=result_type,
+                                                       output_type=output_type,
+                                                       filter_steps=filter_steps):
+        if (isinstance(item, ResponseIntermediateStep) or isinstance(item, ResponsePayloadOutput)):
+            yield item.get_stream_data()
+        else:
+            raise ValueError("Unexpected item type in stream. Expected ChatResponseSerializable, got: " +
+                             str(type(item)))

nat/front_ends/fastapi/step_adaptor.py

@@ -0,0 +1,319 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import logging
+from functools import reduce
+from textwrap import dedent
+
+from nat.data_models.api_server import ResponseIntermediateStep
+from nat.data_models.api_server import ResponseSerializable
+from nat.data_models.intermediate_step import IntermediateStep
+from nat.data_models.intermediate_step import IntermediateStepCategory
+from nat.data_models.intermediate_step import IntermediateStepPayload
+from nat.data_models.intermediate_step import IntermediateStepType
+from nat.data_models.invocation_node import InvocationNode
+from nat.data_models.step_adaptor import StepAdaptorConfig
+from nat.data_models.step_adaptor import StepAdaptorMode
+from nat.utils.type_utils import is_valid_json
+
+logger = logging.getLogger(__name__)
+
+
+class StepAdaptor:
+
+    def __init__(self, config: StepAdaptorConfig):
+
+        self._history: list[IntermediateStep] = []
+        self.config = config
+
+    def _step_matches_filter(self, step: IntermediateStep, config: StepAdaptorConfig) -> bool:
+        """
+        Returns True if this intermediate step should be included (based on the config.mode).
+        """
+
+        if config.mode == StepAdaptorMode.OFF:
+            return False
+
+        if config.mode == StepAdaptorMode.DEFAULT:
+            # default existing behavior: show LLM events + TOOL_END + FUNCTION events
+            if step.event_category == IntermediateStepCategory.LLM:
+                return True
+            if step.event_category == IntermediateStepCategory.TOOL:
+                return True
+            if step.event_category == IntermediateStepCategory.FUNCTION:
+                return True
+            return False
+
+        if config.mode == StepAdaptorMode.CUSTOM:
+            # pass only what the user explicitly listed
+            return step.event_type in config.custom_event_types
+
+        return False
+
+    def _handle_llm(self, step: IntermediateStepPayload, ancestry: InvocationNode) -> ResponseSerializable | None:
+        input_str: str | None = None
+        output_str: str | None = None
+
+        # Find the start in the history with matching run_id
+        start_step = next(
+            (x for x in self._history if x.event_type == IntermediateStepType.LLM_START and x.UUID == step.UUID), None)
+
+        if not start_step:
+            # If we don't have a start step, we can't do anything
+            return None
+
+        input_str = str(start_step.data.input)
+
+        if step.event_type == IntermediateStepType.LLM_NEW_TOKEN:
+
+            # Find all of the previous LLM chunks and concatenate them
+            output_str = reduce(
+                lambda x, y: x + y,
+                (str(x.data.chunk)
+                 for x in self._history if x.event_type == IntermediateStepType.LLM_NEW_TOKEN and x.UUID == step.UUID),
+                "")
+
+        elif step.event_type == IntermediateStepType.LLM_END:
+            output_str = str(step.data.output)
+
+        if not input_str and not output_str:
+            return None
+
+        escaped_input = html.escape(input_str, quote=False)
+
+        # Dont use f-strings here because the payload is markdown and screws up the dedent
+        payload = dedent("""
+        **Input:**
+        ```python
+        {input_value}
+        ```
+        """).strip("\n").format(input_value=escaped_input)
+
+        if (output_str):
+            escaped_output = html.escape(output_str, quote=False) if output_str else ""
+
+            # Dont use f-strings here because the payload is markdown and screws up the dedent
+            payload = dedent("""
+            {payload}
+
+            **Output:**
+            {output_value}
+            """).strip("\n").format(payload=payload, output_value=escaped_output)
+
+        event = ResponseIntermediateStep(id=step.UUID,
+                                         name=step.name or "",
+                                         payload=payload,
+                                         parent_id=ancestry.function_id)
+
+        return event
+
+    def _handle_tool(self, step: IntermediateStepPayload, ancestry: InvocationNode) -> ResponseSerializable | None:
+        """
+        Handles both TOOL_START and TOOL_END events
+        """
+        input_str: str | None = None
+        output_str: str | None = None
+
+        # Find the start in the history with matching run_id
+        start_step = next(
+            (x for x in self._history if x.event_type == IntermediateStepType.TOOL_START and x.UUID == step.UUID), None)
+
+        if not start_step:
+            # If we don't have a start step, we can't do anything
+            return None
+
+        input_str = str(start_step.data.input)
+
+        if step.event_type == IntermediateStepType.TOOL_END:
+            output_str = str(step.data.output)
+
+        if not input_str and not output_str:
+            return None
+
+        escaped_input = html.escape(input_str, quote=False)
+        format_input_type = "json" if is_valid_json(escaped_input) else "python"
+
+        # Dont use f-strings here because the payload is markdown and screws up the dedent
+        payload = dedent("""
+        **Input:**
+        ```{format_input_type}
+        {input_value}
+        ```
+        """).strip("\n").format(input_value=escaped_input, format_input_type=format_input_type)
+
+        if output_str:
+            escaped_output = html.escape(output_str, quote=False)
+            format_output_type = "json" if is_valid_json(escaped_output) else "python"
+
+            # Dont use f-strings here because the payload is markdown and screws up the dedent
+            payload = dedent("""
+            {payload}
+
+            **Output:**
+            ```{format_output_type}
+            {output_value}
+            ```
+            """).strip("\n").format(payload=payload, output_value=escaped_output, format_output_type=format_output_type)
+
+        event = ResponseIntermediateStep(id=step.UUID,
+                                         name=f"Tool: {step.name}",
+                                         payload=payload,
+                                         parent_id=ancestry.function_id)
+
+        return event
+
+    def _handle_function(self, step: IntermediateStepPayload, ancestry: InvocationNode) -> ResponseSerializable | None:
+        """
+        Handles the FUNCTION_START and FUNCTION_END events
+        """
+        input_str: str | None = None
+        output_str: str | None = None
+
+        if step.event_type == IntermediateStepType.FUNCTION_START:
+            # For function start events, display input data
+            if step.data and hasattr(step.data, 'input'):
+                input_str = str(step.data.input)
+            elif step.data:
+                input_str = str(step.data)
+
+            if not input_str:
+                return None
+
+            escaped_input = html.escape(input_str, quote=False)
+            format_input_type = "json" if is_valid_json(escaped_input) else "python"
+
+            # Create payload for function start
+            payload_str = dedent("""
+            **Function Input:**
+            ```{format_input_type}
+            {input_value}
+            ```
+            """).strip("\n").format(input_value=escaped_input, format_input_type=format_input_type)
+
+            event = ResponseIntermediateStep(id=step.UUID,
+                                             name=f"Function Start: {step.name}",
+                                             payload=payload_str,
+                                             parent_id=ancestry.parent_id)
+            return event
+
+        if step.event_type == IntermediateStepType.FUNCTION_END:
+            # Find the start event with matching UUID
+            start_step = next(
+                (x
+                 for x in self._history if x.event_type == IntermediateStepType.FUNCTION_START and x.UUID == step.UUID),
+                None)
+
+            # For function end events, display output data
+            if step.data and hasattr(step.data, 'output'):
+                output_str = str(step.data.output)
+            elif step.data:
+                output_str = str(step.data)
+
+            if not output_str:
+                return None
+
+            escaped_output = html.escape(output_str, quote=False)
+            format_output_type = "json" if is_valid_json(escaped_output) else "python"
+
+            # Get input from start step if available
+            input_payload = ""
+            if start_step and start_step.data:
+                if hasattr(start_step.data, 'input'):
+                    input_str = str(start_step.data.input)
+                else:
+                    input_str = str(start_step.data)
+
+                if input_str:
+                    escaped_input = html.escape(input_str, quote=False)
+                    format_input_type = "json" if is_valid_json(escaped_input) else "python"
+                    input_payload = dedent("""
+                    **Function Input:**
+                    ```{format_input_type}
+                    {input_value}
+                    ```
+                    """).strip("\n").format(input_value=escaped_input, format_input_type=format_input_type)
+
+            # Create payload for function end
+            payload_str = dedent("""
+            {input_payload}**Function Output:**
+            ```{format_output_type}
+            {output_value}
+            ```
+            """).strip("\n").format(input_payload=input_payload,
+                                    output_value=escaped_output,
+                                    format_output_type=format_output_type)
+
+            event = ResponseIntermediateStep(id=step.UUID,
+                                             name=f"Function Complete: {step.name}",
+                                             payload=payload_str,
+                                             parent_id=ancestry.parent_id)
+            return event
+
+        return None
+
+    def _handle_custom(self, payload: IntermediateStepPayload, ancestry: InvocationNode) -> ResponseSerializable | None:
+        """
+        Handles the CUSTOM event
+        """
+        escaped_payload = html.escape(str(payload), quote=False)
+        escaped_payload = escaped_payload.replace("\n", "")
+
+        # Attempt to determine type
+        format_type = "json" if is_valid_json(escaped_payload) else "python"
+
+        # Don't use f-strings here because the payload is markdown and screws up the dedent
+        payload_str = dedent("""
+        ```{format_type}
+        {payload}
+        ```
+        """).strip("\n").format(payload=escaped_payload, format_type=format_type)
+
+        # Return the event
+        event = ResponseIntermediateStep(id=payload.UUID,
+                                         name=f"{payload.event_type}",
+                                         payload=payload_str,
+                                         parent_id=ancestry.function_id)
+
+        return event
+
+    def process(self, step: IntermediateStep) -> ResponseSerializable | None:  # pylint: disable=R1710
+
+        # Track the chunk
+        self._history.append(step)
+        payload = step.payload
+        ancestry = step.function_ancestry
+
+        if not self._step_matches_filter(step, self.config):
+            return None
+
+        try:
+
+            if step.event_category == IntermediateStepCategory.LLM:
+                return self._handle_llm(payload, ancestry)
+
+            if step.event_category == IntermediateStepCategory.TOOL:
+                return self._handle_tool(payload, ancestry)
+
+            if step.event_category == IntermediateStepCategory.FUNCTION:
+                return self._handle_function(payload, ancestry)
+
+            if step.event_category == IntermediateStepCategory.CUSTOM:
+                return self._handle_custom(payload, ancestry)
+
+        except Exception as e:
+            logger.error("Error processing intermediate step: %s", e, exc_info=True)
+
+        return None

nat/front_ends/mcp/__init__.py

@@ -0,0 +1,14 @@
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

nat/front_ends/mcp/mcp_front_end_config.py

@@ -0,0 +1,36 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from pydantic import Field
+
+from nat.data_models.front_end import FrontEndBaseConfig
+
+
+class MCPFrontEndConfig(FrontEndBaseConfig, name="mcp"):
+    """MCP front end configuration.
+
+    A simple MCP (Modular Communication Protocol) front end for NeMo Agent toolkit.
+    """
+
+    name: str = Field(default="NeMo Agent Toolkit MCP",
+                      description="Name of the MCP server (default: NeMo Agent Toolkit MCP)")
+    host: str = Field(default="localhost", description="Host to bind the server to (default: localhost)")
+    port: int = Field(default=9901, description="Port to bind the server to (default: 9901)", ge=0, le=65535)
+    debug: bool = Field(default=False, description="Enable debug mode (default: False)")
+    log_level: str = Field(default="INFO", description="Log level for the MCP server (default: INFO)")
+    tool_names: list[str] = Field(default_factory=list,
+                                  description="The list of tools MCP server will expose (default: all tools)")
+    runner_class: str | None = Field(
+        default=None, description="Custom worker class for handling MCP routes (default: built-in worker)")

nat/front_ends/mcp/mcp_front_end_plugin.py

@@ -0,0 +1,81 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import typing
+
+from nat.builder.front_end import FrontEndBase
+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
+
+logger = logging.getLogger(__name__)
+
+
+class MCPFrontEndPlugin(FrontEndBase[MCPFrontEndConfig]):
+    """MCP front end plugin implementation."""
+
+    def get_worker_class(self) -> type[MCPFrontEndPluginWorkerBase]:
+        """Get the worker class for handling MCP routes."""
+        from nat.front_ends.mcp.mcp_front_end_plugin_worker import MCPFrontEndPluginWorker
+
+        return MCPFrontEndPluginWorker
+
+    @typing.final
+    def get_worker_class_name(self) -> str:
+        """Get the worker class name from configuration or default."""
+        if self.front_end_config.runner_class:
+            return self.front_end_config.runner_class
+
+        worker_class = self.get_worker_class()
+        return f"{worker_class.__module__}.{worker_class.__qualname__}"
+
+    def _get_worker_instance(self) -> MCPFrontEndPluginWorkerBase:
+        """Get an instance of the worker class."""
+        # Import the worker class dynamically if specified in config
+        if self.front_end_config.runner_class:
+            module_name, class_name = self.front_end_config.runner_class.rsplit(".", 1)
+            import importlib
+            module = importlib.import_module(module_name)
+            worker_class = getattr(module, class_name)
+        else:
+            worker_class = self.get_worker_class()
+
+        return worker_class(self.full_config)
+
+    async def run(self) -> None:
+        """Run the MCP server."""
+        # Import FastMCP
+        from mcp.server.fastmcp import FastMCP
+
+        # 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 and set up routes
+        worker = self._get_worker_instance()
+
+        # 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
+            await mcp.run_sse_async()