Configuring external tools with MCP servers

Add, manage, and configure MCP servers for Delegate and Cartographer, including custom servers, settings.json, and MCP Apps.

MCP is an open protocol for connecting AI assistants to external data sources and tools. Unlike Integration Service connectors, MCP servers can be remote (hosted elsewhere, accessed over HTTP) or local (a program running on your machine).

Most MCP servers don't need manual configuration — see MCP servers reference for the Orchestrator-backed catalog and the third-party marketplace, both of which add a server with one click.

Adding a custom MCP server

For a server that isn't in the catalog or marketplace, add it manually from Add-ons & Integrations → MCP servers.

Remote MCP servers

Remote MCP servers are hosted elsewhere and accessed over HTTP:

1. Select Add MCP server → Remote.
2. Enter the server details:
   • URL: the endpoint address (for example, https://api.example.com/mcp)
   • Title: a friendly name (for example, "Company API")
   • Headers (optional): authentication headers, if required
3. Select Save.
4. Toggle Enabled to activate.

Example configuration:

{
  "type": "remote",
  "url": "https://internal-api.company.com/mcp",
  "enabled": true,
  "title": "Company Data API",
  "headers": {
    "Authorization": "<your-access-token>"
  }
}

Local MCP servers

Local MCP servers run as processes on your machine:

1. Select Add MCP server → Local.
2. Enter the server details:
   • Title: a friendly name
   • Command: the executable to run (for example, node, python, /path/to/server)
   • Arguments: command-line arguments
   • Working directory: where the command runs
   • Environment variables: key-value pairs for the process
3. Select Save.
4. Toggle Enabled to activate.

Example configurations

Filesystem MCP server:

{
  "type": "local",
  "enabled": true,
  "serverId": "filesystem-mcp",
  "title": "Local Filesystem Access",
  "config": {
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/username/Documents"],
    "env": {},
    "cwd": "/Users/username"
  }
}

GitHub MCP server:

{
  "type": "local",
  "enabled": true,
  "serverId": "github-mcp",
  "title": "GitHub Integration",
  "config": {
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-github"],
    "env": {
      "GITHUB_PERSONAL_ACCESS_TOKEN": "<your-github-token>"
    },
    "cwd": "/Users/username"
  }
}

Managing MCP servers

From Add-ons & Integrations → MCP servers, you can toggle a configured server on or off, edit its configuration, or remove it. Disabling a server stops the agent from using its tools, but doesn't stop a local server process that's already running — you may need to stop that process yourself.

How the agent discovers MCP tools

When an MCP server is enabled, the agent connects to it, the server reports its available tools, and those tools become available in conversation — no manual registration needed. You can reference them in natural language, for example: "Use the filesystem tool to list my Documents folder."

MCP Apps: inline interactive interfaces

MCP Apps is an extension to the Model Context Protocol that lets a tool return an interactive interface — a form, a dashboard, a chart, a data table — instead of only text. When a connected MCP server's tool supports it, Delegate renders that interface inline in the conversation, in a sandboxed frame: the interface can call the server's own tools and receive updates, but can't access anything outside its own frame. Forms arrive pre-filled with sensible values, so you only change what matters instead of describing every field in chat.

Whether a given MCP server offers this depends on that server — not every server implements the extension.

Administrators can govern MCP servers at the policy level — see MCP Servers in Governance with Automation Ops.

Advanced: connection management via settings.json

Power users can manage MCP server connections directly in settings.json — the same mcpServers list the UI reads and writes:

{
  "mcpServers": [
    {
      "type": "remote",
      "url": "https://api.example.com/mcp",
      "enabled": true,
      "title": "Example API",
      "headers": {
        "Authorization": "<your-access-token>"
      }
    },
    {
      "type": "local",
      "enabled": true,
      "serverId": "my-local-server",
      "title": "Local Tools",
      "config": {
        "command": "python",
        "args": ["/path/to/server.py"],
        "env": { "API_KEY": "<your-api-key>" },
        "cwd": "/path/to/working/dir"
      }
    }
  ]
}

Edit settings.json only if you're comfortable with JSON — invalid syntax prevents the app from loading. Always back up before editing. See Settings file locations.

For security considerations and data flow information for connections, see Local security configuration.

Next steps

• MCP servers reference — The Orchestrator catalog and third-party marketplace
• Connecting to popular business applications — For managed SaaS integrations
• Repeating work with Routines — Save workflows that use your connected tools
• Sharing skills with your team — Find pre-built skills for your connected services