What people run on it
Each of these is a single chat message away once a device is connected. The prompts below are the literal shape of how it's used — not marketing paraphrase.
Reach the machine you left behind
The common case, and the one most people start with: a computer that is powered on somewhere else, with the work already open on it. No file sync, no copies, no getting the environment right on a second device — the agent operates the one that already has everything.
my laptop's at home and I'm not — open the doc I was editing and read me the last section find the invoice PDF I saved last week somewhere in downloads and put it where I can grab it the thing I left running looks wedged. take a screenshot, tell me what's on screen, and kill it if it's hungEvent-driven monitoring
Anything that can send an HTTP request can wake an agent. Detectors, sensors, cron jobs, CI — they post an event, the event routes to an agent, the agent investigates and writes what it found to the activity feed. Alerts stay separate from chatter.
when the camera script posts a motion event after 10pm, grab a snapshot from the office machine and put it in the feed as an alertAmbient and personal
The same machinery at household scale: routines that glance at things for you, voice when hands are busy, notes with photos pulled straight off a device.
photograph the water meter with the workshop webcam every morning and log the reading — flag it if the daily jump is ever double the average read me the feed summary — car mode, keep listeningProvision a GPU from the conversation
Run the job and watch it without babysitting a terminal. Credits are metered per minute; storage persists between sessions so checkpoints survive.
spin up an A10 with my data volume, pull the repo, and start the training run — pin a card with the loss curve and ping the feed if it plateaus every night at 2am, run the parameter sweep on the lab workstation and post a summary table of the resultsWorks for literature monitoring too: a scheduled agent that searches, filters against what it has already seen, and posts only what's new.
Once there's more than one
Put the devices in a group and address them as one. The agent resolves which device a task belongs to, or fans out across all of them and reports per-device. The same conversation, reaching further.
check all six kiosks: is the display app running and on the right screen? note anything odd to the feed, don't wake me unless one is down on every machine in the render group, show me disk space and the last backup time as one tableWithout an account
The interaction model can be felt before signing up: the front page is a live build agent, the same loop with browser-only tools.