Control Center

The task

Several projects in parallel, two LLM subscriptions and a server full of deployments. I needed to see everything in one place and not hit a limit at the worst moment.

What I built

• A task queue with leases, heartbeats and cancellation: agents on a Mac and a VPS pick up tasks by role, and a cancelled task is killed along with its child processes
• A pipeline: a strong model plans and accepts the work, a local one in Ollama does it, plain code checks the result, and a failure escalates
• Limits are read straight from the subscription, the burn rate gives a forecast like “runs out at 02:10”, and heavy tasks wait for the window reset or for the night on their own
• A scheduler with launch conditions, daily and weekly reports, a Sites section with domain checks and on / off switches
• An LLM agent console in the browser and on the phone, a PWA, web push without libraries, a Telegram bot with a morning digest