MCP access

Let Cursor, Claude or Codex read your job configs, metrics and logs.

MCP exposes job data to an AI client so you can ask about a run where you are already working, instead of copying log excerpts into a chat window.

MCP connection details and per-client configuration.

What you see

Region	What it shows
Connection	The MCP endpoint, over Streamable HTTP
Configure your client	Ready-to-paste configuration for Cursor, Claude Desktop, Claude Code or Codex
Available tools	Which tools your account permits, as a set of switches

Connecting

1. Pick your client

   The page generates the right configuration for each: Cursor writes ~/.cursor/mcp.json, Claude Desktop goes through mcp-remote, Claude Code and Codex are added from the command line.

2. Save the configuration, then click Connect

   Authentication is OAuth 2.1 with PKCE. The first connection opens a browser to authorise; after that the client holds a token.

3. Ask about a job

   Tools cover job configuration, hyperparameters, metrics and logs.

Tool permissions apply to your MCP account globally — every client you connect shares them. Turning a tool off here turns it off everywhere.

What is exposed, and what is not

Read tools are the default. Write tools are off unless you enable them, and a deployment can restrict MCP to administrators entirely with FORGE_PREMIUM_PUBLIC.

An MCP client sees what your account can see — no more. It is not a way around dataset governance: a protected dataset’s records are not readable through the console, so they are not readable here either.

A deployment running with authentication disabled skips OAuth entirely. That is for local development, and the page says so when it applies.