Skip to content

Setting Up the DataForge MCP Server

DataForge includes a Talos-hosted MCP (Model Context Protocol) server that lets AI clients such as Claude Code, Claude, and Codex use DataForge tools with your DataForge credentials. Every tool available to the Talos AI assistant is available through it: reading configuration, creating and editing sources and outputs, running pulls, and checking processing status — governed by the same project roles as the UI.

The server speaks streamable HTTP with standard OAuth 2.1 authorization, so any MCP client that supports remote servers can connect. No API keys or tokens are created or pasted anywhere — each user signs in interactively with their own DataForge credentials.

Find your MCP URL

  1. In DataForge, open the System menu. This opens the /service-configs page.
  2. Find the api-url property and copy its value.
  3. Build the MCP URL as https://{api-url}/talos/mcp, replacing {api-url} with the value you copied.

Claude Code

Add the server from a terminal:

claude mcp add --transport http dataforge https://{api-url}/talos/mcp

By default this configures the server for you in the current project. Add --scope user to make it available in all your projects, or --scope project to write a .mcp.json file that can be committed so the whole team gets the server. A committed .mcp.json looks like this:

{
  "mcpServers": {
    "dataforge": {
      "type": "http",
      "url": "https://{api-url}/talos/mcp"
    }
  }
}

Then authenticate: inside Claude Code, run /mcp, select dataforge, and choose Authenticate. A browser window opens — sign in with your normal DataForge credentials and approve the connection.

Verify: run /mcp again and confirm the server shows as connected, then ask Claude Code to list the available DataForge tools.

Claude (web and desktop)

In Claude on the web or the desktop app, the server is added as a custom connector:

  1. Open Settings → Connectors and choose Add custom connector.
  2. Name it DataForge and enter the MCP URL: https://{api-url}/talos/mcp.
  3. Click Connect, sign in with your DataForge credentials, and approve the connection.

Team and Enterprise plans may require an admin to enable custom connectors for the organization.

Codex

Codex configures MCP servers in ~/.codex/config.toml. Add:

[mcp_servers.dataforge]
url = "https://{api-url}/talos/mcp"

OAuth is the default authentication for remote servers, so no token configuration is needed. Then sign in:

codex mcp login dataforge

A browser window opens — sign in with your DataForge credentials and approve the connection.

Verify: codex mcp list shows the server, and a Codex session can list the DataForge tools.

Other MCP clients

Any client that supports remote MCP servers over streamable HTTP with OAuth authorization can connect the same way: add a server with the URL https://{api-url}/talos/mcp, and complete the sign-in prompt in the browser when the client first connects.

After connecting

  • You will see 11 tools, not 130. This is expected. The server presents a compact 11-tool surface: 5 discovery meta-tools, 3 project-session tools, and 3 instruction-serving tools, with call_dataforge_tool as the single gateway every other call passes through. Behind that gateway sits the full DataForge tool registry. The surface is kept small on purpose, so an assistant does not spend its context window reading tool definitions it will never call. Use the discovery tools to find the registry tool you need, then dispatch it through the gateway.
  • Start with the instructions tool. The server publishes its own usage guide — have your assistant call dataforge_agents_instructions at the start of a session. It returns the full agent guide and the DataForge object taxonomy, and the assistant will make far better tool choices with it loaded.
  • Access follows your user. Tool calls run as you, against the projects your DataForge roles allow — the same permissions as the UI. Use select_project to point the session at the project you want to work in.
  • For unattended automation, use the External API instead. MCP authentication is an interactive browser sign-in, so it suits people working alongside an assistant. Scheduled jobs and CI/CD should use machine credentials and the M2M routes, such as the automated production Pull.