It is set up and on your desk. Here is what every screen means, what happens when you tap, which agents you can actually stop from the glass, and the things it deliberately will not do.
It is a clock for your agents. You do not open a clock and you do not manage a clock. You glance, you know, you go back to work. The device is calm on purpose, and only gets loud for the two things that genuinely need a human: an agent that is stuck or burning money, and an agent that is blocked waiting on your decision.
If you have not set it up yet, start at device setup. If you bought a bare board, flash it first.
The resting screen answers four questions without you asking any of them:
If you use Claude, the device also carries the 5 hour and 7 day limit meters with the reset countdown, so you can see how much of your window is left before you start something large.
The totals are fleet-wide. Pairing links the device to your account, not to one machine, so spend, tokens, live agents, alerts and approvals are summed across every machine reporting to that account. The live event feed is the one exception: each machine seals its events with its own key, so the device cycles through your machines rather than showing them all in one stream.
| Screen | When you see it | What to do |
|---|---|---|
| All clear | Everything is working and nothing needs you. | Nothing. This is the state it should be in most of the day. |
| Glance | Agents are active. Today's spend, per-runtime tiles, live counts. | Read it and move on, or tap a runtime to go deeper. |
| Attention | An agent is looping or has made no progress for a while. | Decide whether to let it run or stop it. The screen names the runtime and what it has been repeating. |
| Waiting on you | An agent paused and asked for permission to do something. | Read the exact request and tap Approve or Deny. The agent unblocks immediately. |
| Drill-down | You tapped a runtime on the glance. | See that one agent's spend today, sessions, tools and recent activity. In preview; it is rolling out over the air. |
| Offline | The device cannot reach your account, or nothing has reported recently. | Usually nothing. It reconnects on its own and says how long ago it last had fresh data, rather than showing you a stale number as if it were current. |
| Pairing | The device is not linked to an account yet, or you cleared its pairing. | Scan the QR, or enter the short code at clawmetry.com/activate. |
When an agent pauses to ask permission, the device shows the exact request, the runtime that asked, and how long it has been waiting. You tap Approve or Deny on the glass and the agent unblocks on the spot. No laptop, no hunting for the right terminal tab.
Approve and Deny come from runtimes that pause and wait for a decision, which today means OpenClaw and NVIDIA NemoClaw. Every decision runs through the same approvals path as the ClawMetry web app, and every one is logged.
An approval is an agent asking. This is you deciding without being asked. From the glass you can:
These work on Claude Code, Codex, Goose, opencode and Aider, alongside OpenClaw and NemoClaw. Cursor is the exception: the editor runs all of its conversations inside one process, so there is no single session to stop, and the device says so rather than offering a button that does nothing.
The device never acts on its own. Watching is automatic; stopping is always a deliberate human tap. The device surfaces the runaway and makes the call one tap away, but the call is yours, and every action is recorded.
Your agents' conversations and activity are encrypted on your machine before they leave it. The device holds your key and decrypts on the screen itself, so the cloud stores ciphertext it cannot read. Simple totals, meaning cost, token counts and health status, are sent securely in readable form so the glance is instant and your dashboard agrees with your desk.
The key is generated on the device during setup and never touches our servers in readable form. Revoking the device from your account cuts its access immediately.
Worth saying plainly, so nothing about it surprises you later:
Yes. Each device pairs to an account and shows that account's fleet, so a second device on the same account shows the same glance. Pair each one separately from its own screen.
The device switches to its Offline screen and tells you when it last had fresh data, then reconnects on its own. It deliberately does not keep showing the last number as though it were current; a status display that lies quietly is worse than one that admits it is stale.
The device is $49 once and there is no device subscription, ever. On the software side, OpenClaw and NVIDIA NemoClaw are on the free plan; watching Claude Code, Codex, Cursor and the other runtimes needs Starter at $9 per node per month, with a free trial and no card. See pricing.
Every runtime ClawMetry supports, including OpenClaw, NVIDIA NemoClaw, Claude Code, Codex, Cursor, Aider, Goose, opencode, Qwen Code, Hermes, PicoClaw, NanoClaw, Pi, Deep Agents, n8n, Antigravity and GitHub Copilot. One integration covers them all, so the device does not go blind the day you add a new tool.
Not today. The ClawMetry software self-hosts, but the device pairs and reads through ClawMetry Cloud, so if the device is your goal, plan for a cloud account.
