Server Manager

Server Manager is an agent mode for multi-server setups. Instead of exposing every tool from every server at initialization, the agent starts with management tools and connects to individual servers when the task needs them. Use Server Manager when you connect to several servers or when the combined tool list is too large for the model context. For one or two small servers, the normal agent mode is usually simpler.

How it changes initialization

With useServerManager: true, the agent receives management tools first. When the agent connects to a server, the available tool set changes and the agent executor is recreated with the newly loaded tools.

Enable Server Manager

Pass useServerManager: true when constructing the agent. Explicit mode requires a client because Server Manager coordinates configured client servers.

import { MCPClient } from '@mcp-use/client'
import { MCPAgent } from '@mcp-use/agent'
import { ChatOpenAI } from '@langchain/openai'

const client = new MCPClient({
mcpServers: {
playwright: {
command: 'npx',
args: ['@playwright/mcp@latest']
},
filesystem: {
command: 'npx',
args: ['-y', '@modelcontextprotocol/server-filesystem', '/tmp']
}
}
})

const agent = new MCPAgent({
llm: new ChatOpenAI({ model: 'gpt-4o' }),
client,
useServerManager: true
})

const result = await agent.run({
prompt: 'Use the right server to inspect /tmp and summarize what you find.',
maxSteps: 12
})

console.log(result)
await agent.close()

Management tools

The agent uses these management tools before it can call tools from a configured MCP server.

Tool	What it does
`list_mcp_servers`	Lists configured servers and their known capabilities.
`connect_to_mcp_server`	Connects to one server and loads its tools.
`get_active_mcp_server`	Reports the currently active server.
`disconnect_from_mcp_server`	Disconnects the active server and removes its tools.
`add_mcp_server_from_config`	Adds a server from configuration during the run.

Prompt for server discovery

Ask the agent to inspect available servers before choosing one. This helps when server names are not obvious from the user request.

import { MCPClient } from '@mcp-use/client'
import { MCPAgent } from '@mcp-use/agent'
import { ChatOpenAI } from '@langchain/openai'

async function demoServerManager() {
  const client = new MCPClient({
    mcpServers: {
      web: { command: 'npx', args: ['@playwright/mcp@latest'] },
      files: {
        command: 'npx',
        args: ['-y', '@modelcontextprotocol/server-filesystem', '/tmp']
      }
    }
  })

  const agent = new MCPAgent({
    llm: new ChatOpenAI({ model: 'gpt-4o' }),
    client,
    useServerManager: true,
    verbose: true
  })

  const result = await agent.run({
    prompt: `
      First list the available MCP servers.
      Then choose the right server to inspect /tmp.
      Summarize what you found and mention which server you connected to.
    `,
    maxSteps: 12
  })

  console.log(result)
  await agent.close()
}

demoServerManager().catch(console.error)

When not to use it

Do not enable Server Manager just to connect to a single server. It adds an extra discovery step and makes the model choose a server before it can use tools. For small setups, expose all tools directly and rely on disallowedTools, exposeResourcesAsTools, and exposePromptsAsTools to control scope.

Get Started

MCP Server

MCP Apps

MCP Client

MCP Agent

Tooling

How it changes initialization

Enable Server Manager

Management tools

Prompt for server discovery

When not to use it

​How it changes initialization

​Enable Server Manager

​Management tools

​Prompt for server discovery

​When not to use it

How it changes initialization

Enable Server Manager

Management tools

Prompt for server discovery

When not to use it