Asking an AI assistant "did last night's production build succeed" is only useful if the assistant can actually check, rather than guess from whatever context happens to be in the conversation. That's what the MCP server is for.
Forewarden runs a Model Context Protocol server at /api/mcp, the same standard Claude and other MCP-aware clients use to call tools directly. Point a client at it with an API token and it gets access to a defined set of tools: list projects and models, read a model's lineage and health, trigger and check on job runs, run a read-only SQL query, and query a metric, using the same permissions your token already has.
Why permission scoping matters here specifically
An AI assistant with API access to your data platform is a real permissions boundary, not a convenience feature. Every MCP call runs with the token's role: a personal token acts as its owner and a service token as the role it was given, so a Viewer can read lineage and health but cannot trigger a run, the same as a person with that role couldn't from the browser. There's no separate, more permissive path for AI tools than for a person clicking the same button.