Strands MCP Server

Instructions

Implementation Reference

• strands_mcp_server/mcp_server.py:325-371 (handler)

  Handler logic within the MCP call_tool function that creates a fresh agent instance, invokes it with the provided prompt, and returns the response as TextContent.

  if name == "invoke_agent" and expose_agent:
      prompt = arguments.get("prompt")
      if not prompt:
          return [
              types.TextContent(
                  type="text",
                  text="❌ Error: 'prompt' parameter is required",
              )
          ]
  
      logger.debug(f"Invoking agent with prompt: {prompt[:100]}...")
  
      # Get the parent agent's configuration
      # Access tools directly from registry dictionary
      tools_for_invocation = [
          agent.tool_registry.registry[tool_name]
          for tool_name in agent_tools.keys()
          if tool_name in agent.tool_registry.registry
      ]
  
      # Prepare extra kwargs for observability and callbacks
      extra_kwargs = {}
      if hasattr(agent, "callback_handler") and agent.callback_handler:
          extra_kwargs["callback_handler"] = agent.callback_handler
  
      # Create fresh agent with same configuration but clean message history
      # Inherits: model, tools, trace_attributes, callback_handler
      fresh_agent = Agent(
          name=f"{agent.name}-invocation",
          model=agent.model,
          messages=[],  # Empty message history (clean state)
          tools=tools_for_invocation,
          system_prompt=agent.system_prompt if hasattr(agent, "system_prompt") else None,
          trace_attributes=agent.trace_attributes if hasattr(agent, "trace_attributes") else {},
          **extra_kwargs,
      )
  
      # Call the fresh agent
      result = fresh_agent(prompt)
  
      # Extract text response from agent result
      response_text = str(result)
  
      logger.debug(f"Agent invocation complete, response length: {len(response_text)}")
  
      return [types.TextContent(type="text", text=response_text)]

• strands_mcp_server/mcp_server.py:272-290 (schema)

  Defines the MCP Tool schema for 'invoke_agent', including name, description, and input schema requiring a 'prompt' parameter.

  agent_invoke_tool = types.Tool(
      name="invoke_agent",
      description=(
          f"Invoke the full {agent.name} agent with a natural language prompt. "
          "Use this for complex queries that require reasoning across multiple tools "
          "or when you need a conversational response from the agent."
      ),
      inputSchema={
          "type": "object",
          "properties": {
              "prompt": {
                  "type": "string",
                  "description": "The prompt or query to send to the agent",
              }
          },
          "required": ["prompt"],
      },
  )
  mcp_tools.append(agent_invoke_tool)

• strands_mcp_server/mcp_server.py:298-308 (registration)

  Registers the list_tools handler that includes the 'invoke_agent' tool in the returned list of available tools when expose_agent is True.

  @server.list_tools()
  async def list_tools() -> list[types.Tool]:
      """Return list of available MCP tools.
  
      This handler is called when MCP clients request the available tools.
      It returns the pre-built list of MCP Tool objects converted from
      Strands agent tools.
      """
      logger.debug(f"list_tools called, returning {len(mcp_tools)} tools")
      return mcp_tools