PocketOps: an AI server operator in your pocket

The problem

Server administration assumes a laptop and a terminal. But incidents do not schedule themselves around desk time: a container dies while you are out, a disk fills up on a weekend, a deploy needs a nudge from a train. SSH clients on phones technically exist, and typing docker compose logs --tail 200 on a glass keyboard is its own punishment.

What we built

A mobile-first operations assistant whose brain is a headless coding agent running on the server itself. You send a plain-language request from your phone: “the finance app looks down, fix it”. The agent inspects containers, reads logs, forms a hypothesis, applies the fix, verifies the service is healthy, and reports what it did, in sentences, not log dumps.

Design choices that make it trustworthy rather than terrifying:

• The agent is the interface, not an oracle. It executes real commands with real tool access, so answers come from the actual system state, never from a model’s imagination of it.
• Subscription-auth brain. The agent runs on existing subscription credentials rather than metered API keys sitting in a server env file, cutting both cost and key-leak surface.
• Least privilege. The agent operates as a constrained user with access scoped to the app directories and Docker, with dangerous operations gated behind explicit confirmation.
• Async by nature. Long operations stream progress; the phone can lock, the job finishes, the summary arrives.

Architecture

A FastAPI orchestration layer queues tasks for a headless Claude Code process on the host, with a mobile-optimized React front end and Telegram delivery for notifications. Everything runs inside the same reverse-proxy and 2FA perimeter as the rest of the fleet.

Outcome

Routine incidents stopped requiring a laptop. The pattern generalizes: any team running self-hosted infrastructure can bolt a constrained, auditable AI operator onto it and turn “check the logs when you get home” into a two-minute phone interaction.