Strands MCP Server: Bidirectional MCP Integration 🔌

Expose your Strands agent as an MCP server, or connect to external MCP servers.

Bidirectional MCP integration with HTTP/stdio transports, Claude Desktop support, and full agent invocation.

Installation

pip install strands-mcp-server

Overview

graph LR
    subgraph "Your Strands Agent"
        A[Tools: calculator, shell, etc.]
        B[mcp_server tool]
        C[mcp_client tool]
    end

    subgraph "Server Mode"
        B -->|HTTP/stdio| D[MCP Protocol]
        D --> E[Claude Desktop]
        D --> F[Other Agents]
    end

    subgraph "Client Mode"
        H[Remote MCP Servers] -->|HTTP/stdio/SSE| I[MCP Protocol]
        I --> C
    end

    style B fill:#2b6cb0,color:#fff
    style C fill:#38a169,color:#fff

Quick Start

As MCP Server

from strands import Agent
from strands_mcp_server import mcp_server

agent = Agent(tools=[..., mcp_server])
agent("start mcp server on port 8000")

As MCP Client

from strands import Agent
from strands_mcp_server import mcp_client

agent = Agent(tools=[mcp_client])

# Connect to remote server
agent.tool.mcp_client(
    action="connect",
    connection_id="remote",
    transport="http",
    server_url="http://localhost:8000/mcp"
)

# Call remote tool
agent.tool.mcp_client(
    action="call_tool",
    connection_id="remote",
    tool_name="calculator",
    tool_args={"expression": "42 * 89"}
)

Claude Desktop Integration

{
  "mcpServers": {
    "my-agent": {
      "command": "uvx",
      "args": ["strands-mcp-server", "--cwd", "/path/to/project"]
    }
  }
}

API Reference

mcp_server

Parameter	Default	Description
action	required	start, stop, status, list
transport	http	http or stdio
port	8000	Server port
tools	None	Tools to expose (None = all)
expose_agent	True	Include invoke_agent tool
stateless	False	Multi-node ready

mcp_client

Parameter	Description
action	connect, disconnect, list_tools, call_tool
connection_id	Connection identifier
transport	http, stdio, sse
server_url	Server URL
tool_name	Tool to call
tool_args	Tool arguments

invoke_agent

When expose_agent=True, clients can invoke the full agent:

agent.tool.mcp_client(
    action="call_tool",
    connection_id="remote",
    tool_name="invoke_agent",
    tool_args={"prompt": "Calculate 2 + 2 and explain"}
)

CLI

# Start stdio server (for Claude Desktop)
uvx strands-mcp-server

# Start HTTP server
uvx strands-mcp-server --transport http --port 8000

# Proxy mode (stdio ↔ HTTP bridge)
uvx strands-mcp-server --proxy http://remote:8000/mcp

# With custom working directory
uvx strands-mcp-server --cwd /path/to/project

Architecture

sequenceDiagram
    participant CD as Claude Desktop
    participant MCP as MCP Server
    participant Agent as Strands Agent
    participant Tools as Tools

    CD->>MCP: tools/list
    MCP->>Agent: Get tool registry
    Agent->>MCP: Tool definitions
    MCP->>CD: Tool list

    CD->>MCP: tools/call (calculator)
    MCP->>Agent: Route to tool
    Agent->>Tools: Execute
    Tools->>Agent: Result
    Agent->>MCP: Response
    MCP->>CD: Tool result

Modes

HTTP Mode

agent.tool.mcp_server(
    action="start",
    transport="http",
    port=8000
)
# Access at http://localhost:8000/mcp

Stdio Mode

agent.tool.mcp_server(
    action="start",
    transport="stdio"
)
# For Claude Desktop integration

Stateless Mode

For multi-node deployments:

agent.tool.mcp_server(
    action="start",
    stateless=True  # No session state
)

---

Connect anything to anything via MCP. 🔌