Working with Multiple MCP Servers 🪐🪐

One MCP server is great, but real-world workflows often need several specialised back-ends – think scraping the web, querying a database, and editing files, all in one agent. Good news: MCP-Use was built for exactly this.

This tutorial shows you how to configure multiple servers, route tool calls, and keep resource usage in check.

1. Why Multi-Server?

2. Configuration File

Create multi_server_config.json:

{
  "mcpServers": {
    "airbnb": {
      "command": "npx",
      "args": ["-y", "@openbnb/mcp-server-airbnb", "--ignore-robots-txt"]
    },
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"],
      "env": {"DISPLAY": ":1"}
    }
  }
}

Pro-tip: You can mix local processes and remote HTTP/WS endpoints in the same file. See the docs for field definitions.

3. Simple Agent that Uses Both Servers

import asyncio
from dotenv import load_dotenv
from langchain_openai import ChatOpenAI
from mcp_use import MCPClient, MCPAgent

load_dotenv()

async def main():
    client = MCPClient.from_config_file("multi_server_config.json")
    agent = MCPAgent(
        llm=ChatOpenAI(model="gpt-4o"),
        client=client,
        max_steps=40,
        verbose=True
    )

    result = await agent.run(
        "Find a beachfront Airbnb in Barcelona for 2 nights next month, "
        "then open the first listing in a browser and screenshot it."
    )
    print("Final answer: ", result)

    await client.close_all_sessions()

if __name__ == "__main__":
    asyncio.run(main())

By default the agent has access to all tools provided by any server. The LLM decides which to call.

4. Explicit Server Routing

If you want to force a query to a particular server:

await agent.run("Show me Airbnb listings in Tokyo", server_name="airbnb")

This is handy for multi-turn applications where you maintain conversation state yourself.

5. Automatic Server Manager 🤖

For bigger configs it's better to let MCP-Use handle routing:

agent = MCPAgent(
    …,
    use_server_manager=True  # 🔥 auto-detects based on tool name
)

The agent examines every tool call returned by the LLM, picks the correct server, and connects on-demand. Idle servers are disconnected to save RAM.

6. Performance Tips

1. Set sensible max_steps – each step might spin up a new process.
2. Reuse sessions – share a single MCPClient across multiple agents if you're running them in the same process.
3. Limit tools – pass allowed_tools or disallowed_tools to the agent to reduce confusion.

7. Cleanup Checklist ✅

try:
    …
finally:
    await client.close_all_sessions()

This closes WebSockets and kills subprocesses. Forgetting this can leave orphan browsers eating memory.

Congratulations – you now command an army of MCP servers! Combine them creatively and let your LLMs do the heavy lifting.