Skip to main content

Using Dexto CLI to group MCP servers together

Dexto can operate in MCP Tools Mode, where it acts as a local tool aggregation server that groups MCP servers and re-exposes them all under 1 common MCP server.

Unlike the regular MCP server mode where you interact with a Dexto AI agent, this mode provides direct access to the underlying tools without an AI intermediary.

This is useful when you want to:

  • Access tools from multiple MCP servers through a single connection
  • Group tools directly without AI agent processing
  • Create a centralized tool hub for your development environment

How It Works​

In MCP Tools Mode, Dexto:

  1. Connects to multiple MCP servers as configured
  2. Aggregates capabilities discovered at startup from these servers
  3. Exposes them directly as its own local MCP server
  4. Acts as a pass-through for tool execution

Configuration​

For an MCP-only YAML file, run dexto mcp --group-servers --config ./dexto-tools.yml. This avoids agent/model setup. You can create the same configuration using the standalone MCP CLI. The existing agent configuration fallback remains available when --config is omitted.

Step 1: Create a Dexto Configuration File​

Create a dexto-tools.yml configuration file with the MCP servers you want to aggregate:

# dexto-tools.yml
mcpServers:
  filesystem:
    type: stdio
    command: npx
    args:
      - -y
      - "@modelcontextprotocol/server-filesystem"
      - "."
  
  playwright:
    type: stdio
    command: npx
    args:
      - "-y"
      - "@playwright/mcp@latest"
  • You don't need LLM configuration for tools mode
  • Only the mcpServers section is used

Step 2: Setup in Cursor​

Add the following to your .cursor/mcp.json file:

{
  "mcpServers": {
    "dexto-tools": {
      "command": "npx",
      "args": [
        "-y", 
        "dexto", 
        "mcp",
        "--group-servers",
        "-a",
        "path/to/your/dexto-tools.yml"
      ]
    }
  }
}

Or use the default Dexto configuration

{
  "mcpServers": {
    "dexto-tools": {
      "command": "npx",
      "args": [
        "-y", 
        "dexto", 
        "mcp",
        "--group-servers"
      ]
    }
  }
}

Step 3: Restart Cursor​

After adding the configuration, restart Cursor to load the new MCP server.

Connection ownership and protocol behavior​

The grouped server forwards tool schemas and metadata, structured tool results (including isError results), resource metadata and prompt arguments directly through MCP. It does not run an AI agent or grant additional authority: your MCP client and each upstream server own their authorization requirements.

Capabilities are a fixed snapshot of the initial discovery page from each connected upstream. Restart the grouped server after upstream capability changes. Resource templates and discovery pagination are not aggregated. Tools with conflicting names use the Core manager's qualified names, such as filesystem--read_file; duplicate prompt names and ambiguous tool or resource identities fail startup rather than silently selecting a server.

Strict startup requires every enabled configured server to connect. Otherwise, connections marked lenient may fail while successful connections remain available. Shutdown or transport closure releases the grouped server's upstream connections; failed initialization also releases acquired connections. Disabled connections remain disabled even with strict startup.