Chatbots answer. Laptop agents sleep. Hosted Hermes keeps working.

There is a ladder of usefulness in AI tools. Chatbots answer questions. Agents do tasks. But almost every agent you can run today shares one embarrassing dependency: your laptop lid.

The pilot always starts well. You give a coding agent a long task, it works, you watch it. Then the machine sleeps, or the terminal session dies — and the process is gone. Nothing reports it. The scheduled check you set up never fires. Work stops until you sit down again. The agent did not fail at reasoning; it failed at existing.

The failure that matters is boring

Nobody writes postmortems about laptop sleep, because it is not dramatic. But it is the single failure that decides whether an agent is a toy or a worker. A worker keeps working when you leave the room. Watchers keep firing at 3am. A long migration keeps migrating. A CI babysitter is still watching on the fourth retry.

Hosted Hermes is our answer: one always-on agent on a fenced VPS. Not a process on your machine. The run keeps a lease on a hosted box that does not sleep, and you steer it from the browser at thumbgate.app.

Always-on without always-trusted

The moment an agent survives your absence, a second problem appears: what is it allowed to do while you are gone? Our answer is a hard human gate. Automations can draft and run, but money, customer-facing, and production actions pause until you approve or deny them in the browser. Destructive commands, secret leaks, and spend overruns are checked before the tool call executes, not confessed afterwards.

That split — durable execution, gated consequences — is the whole product, and it comes as one plain offer: