Track usage
Two counts matter: what your gadgets did, and how much of your Grok Bot plan the work used. The gateway records the first. Grok Bot shows the second. Neither one measures the other.
On this page
Gadget calls: the gateway log
The service and tunnel route use a request log. Each tool call records the token name, client, device, capability, tool, outcome, duration and bytes. Summarize it:
grok-gadgets-gateway usage
grok-gadgets-gateway usage --token grok-bot --since 2026-10-13
The summary groups calls by token, device, capability and UTC day. Add --json for scripts. Full options: gateway usage guide.
For a live feed, start the gateway with --usage-hook CMD. Your program gets one JSON line per call on its standard input. The line holds no tokens, arguments or device state. A slow or failed program never delays a call.
Credits per account: your relay
A relay with accounts can also keep credits for each account (gateway 0.1.0a20 or later). Commands and online gadget-days use them, and the relay refuses a command at zero. credits report shows what each account spent and checks it against the request log. Credits are a share of your relay, not money and not Grok Bot plan usage. See Credits on your relay.
Grok Bot plan usage: Grok Bot's own pages
- Desktop: Settings → Usage & Billing → Weekly usage limit.
- Phone: Settings → Usage → Weekly usage.
The pages show whole percent, and the allowance resets weekly. Grok Bot measures its own work, not tool calls. No personal usage export or API is documented (research note). The gateway never reads your Grok Bot account.
An estimate per call
Read Weekly usage before and after a session, and note both times. Only calibrate when the usage change is measurable. Then let the gateway work out a rate:
grok-gadgets-gateway usage --token grok-bot \
--since 2026-10-13T09:00 --until 2026-10-13T10:00 \
--calibrate 2 --unit "% weekly usage"
Later summaries add a column labelled estimate. It covers everything Grok Bot did in that window, not only the tool calls.
In the owner's 13 October test, 11 messages and 9 tool calls left the page at 1%. That puts the session under 1 percentage point, as an upper bound (ADR 007). A longer session gives a better rate.
When the reading does not move
Do not calibrate with zero change. Gateway 0.1.0a13 refuses it; 0.1.0a12 accepts zero and can save a misleading zero-rate estimate.
The whole-percent display can hide small usage changes. It does not prove that calls use no allowance.
Gateway #166 tracks the correction.
Hub #271 tracks another day and a useful usage measurement.