Letting engineers query production with AI — without handing out credentials

// the challenge

Engineers at one of our partner companies wanted to use AI coding tools to debug production — query the database, search logs, check metrics — without leaving their editor. The naive way to do that is to paste database credentials and Grafana API keys into everyone's local config.

// our approach

• Hosted MCP gateways inside the platform itself: one for databases, one for logs and metrics (Loki and Prometheus). The AI tool talks to a service we control, never to the database directly.
• Put an authentication proxy in front with Google SSO, so each engineer connects as themselves and no shared secrets ever leave the platform.
• Enforced read-only on the server side: the database backends run from a locked read-only configuration and the observability gateway has all write tools disabled. Nothing a client sends can override it.
• Logged every query with the engineer's identity attached — a full audit trail of who asked what, plus a usage report showing who actually uses the gateways.
• Deployed separate endpoints for sandbox and production, so access to each is granted independently.

// the outcome

Engineers now debug production by asking questions in plain English from their editor, and the answers come from real production data. Nobody holds a database password. Because read-only is enforced on the server, not politely requested in a config file, a bad prompt or a misbehaving tool cannot write to production. And since every query is tied to a person, the client got something it never had with shared credentials: a real audit trail for production access.