AI Usage

Monitors usage, limits and costs of your AI accounts and writes them into read-only ioBroker states. The adapter only reads — it never calls a model, never changes anything at the provider, and never sends your data anywhere.

What it can watch

The three subscription endpoints are the ones those providers' own tools use. They are not officially documented and can change without notice. Claude was tested against a live subscription; ChatGPT and Google are built from verified sources but were never run against a real account — please open an issue if something looks wrong.

Requirements

• Node.js >= 22
• ioBroker js-controller >= 7.2.2
• ioBroker Admin >= 8.0.11 — the adapter reads API keys from the admin's central credential storage instead of asking for them again

Setting it up

1. Install the adapter and open the instance settings.
2. The settings page shows one list: the three subscriptions first, then one row per AI key you stored under Admin → Settings → Credentials.
3. Switch on what you want to watch. Each row has its own warn threshold (10–100 %, default 80 %).
4. For a subscription, the sign-in area opens below its row and walks you through the flow that provider requires. Save first — the sign-in talks to the running instance.
5. After a successful sign-in the account is queried immediately; you do not have to wait for the next cycle.

Options

Accounts are queried in a staggered order, and a provider that answers "too many requests" puts that account into a growing backoff (10 minutes, doubling up to an hour) while the last values stay in place.

The object tree

One device node per account, named the same way for every provider:

ai-usage.0
├─ info.connection            at least one account is delivering data
├─ <account>                  e.g. claude, chatgpt, gemini, <name>-api
│  ├─ info.unreach            the offline marker; drives the icon in the object tree
│  ├─ info.error              why, in plain text; empty while everything works
│  ├─ info.lastUpdate         last successful query
│  ├─ warning                 above the account's warn threshold
│  ├─ limitReached            at 100 %
│  ├─ limits.<window>.percent      utilisation of a limit window
│  ├─ limits.<window>.resetAt      when it resets (empty while no window runs)
│  ├─ limits.<window>.active       whether this window is the limit in force
│  ├─ credits.*               used / limit / remaining / percent, granted / topped up,
│  │                          and whether the balance still covers calls
│  ├─ costs.*                 today / month / total / projected month-end
│  ├─ tokens.*                input and output tokens today
│  └─ models.<model>.*        per-model tokens
└─ total
   ├─ costs.today / month / projectedMonth      summed over all USD accounts
   ├─ maxLimitPercent         the fullest account (limit window or budget)
   ├─ warningsActive          accounts above their threshold
   ├─ limitReached            any account at 100 %
   ├─ accountsReachable       accounts currently delivering
   └─ accounts                accounts you switched on

Datapoints stay once they exist. A provider that leaves a field out for a while does not make its datapoint disappear; time stamps are written empty instead. Only a whole limit window or model that the provider stopped reporting is removed, and switching an account off removes its node completely.

total.costs only sums real money in the same currency — piece counters (request credits, reset vouchers) and other currencies stay out on purpose.

Warnings, and what speaks for an account

Only a plan-wide window can raise an account's warning. A per-model bucket gets its own datapoints but never triggers the alarm: a model you never use can sit at 100 % forever, and an alarm that never clears is worse than no alarm. Google is the exception — it reports no plan-wide window at all, so there the fullest model bucket speaks for the account, and the warning names the model.

The granted budget competes with the windows: an account whose money is nearly spent is just as blocked as one whose time window is full. Whichever side is higher gives the warning its label.

Online status

info.unreach means "this account is not delivering" and drives the connection icon next to the account in the object tree:

Privacy and credentials

• Subscription tokens belong to the adapter alone: they live encrypted in the instance data directory, owner-readable only. The adapter never reads or writes the files of your own tools (~/.codex/auth.json, oauth_creds.json) — those refresh tokens rotate, and two programs refreshing them would sign each other out.
• API keys are read from the admin's central credential storage and never copied.
• The Claude sign-in asks for the profile scope only — the token cannot create API keys or call models.
• The adapter talks to the AI providers and to nobody else.

Troubleshooting

The sign-in button does nothing / the row keeps spinning. Save the settings first, and make sure the instance is running — the sign-in is a conversation with the running adapter.

Google shows an error page after signing in. That is expected and the reason the flow works at all. Copy the whole address out of the address bar and paste it into the field.

"Not signed in" although you signed in. The stored sign-in was rejected by the provider (a revoked or expired refresh token). Sign in again — the row tells you so instead of pretending to be connected.

An OpenAI or Anthropic account delivers nothing. Those reports need an organisation admin key. A personal account without an organisation cannot produce them at all; use the Claude subscription instead.

Claude answers "too many requests". Raise the poll interval. The adapter identifies itself the way Claude's own tooling does and backs off on its own, but a very short interval across several tools can still add up.

Support

Questions, bugs and ideas: https://github.com/krobipd/ioBroker.ai-usage/issues

Changelog

0.13.0 (2026-09-12)

• Fixed: The costs of an Anthropic organisation account were a hundred times too high — the provider counts them in cents, the adapter read them as dollars
• Fixed: The same error was in the adapter-wide cost totals
• Fixed: An account whose values could not be written to the object database kept reporting itself as delivering, with a last-update stamp that went on moving
• Fixed: An answer still waiting on the object database during shutdown could mark accounts online again afterwards, or delete objects
• Fixed: An answer in a shape the adapter does not recognise now counts as a service fault, instead of reading as "this account has nothing"
• Fixed: Limit windows and model channels are no longer deleted when a single answer says nothing about them
• Fixed: Token counters of an organisation account show 0 after UTC midnight instead of keeping yesterday's numbers
• Fixed: Restarting the instance above the warn threshold no longer raises the warning and the notification again
• Fixed: The "limit reached" total now counts a window the provider has closed, the way each account already did
• Fixed: A configured account row the adapter cannot use now says so in the log instead of disappearing
• Improved: The "active window" and DeepSeek "available" flags are written only when they actually change, so their timestamp stops moving on every poll
• Improved: The ChatGPT voucher inventory is fetched about once an hour instead of every poll, halving that account's requests
• Improved: The settings page shows a finished device-code sign-in right away instead of up to half a minute later
• Improved: The provider table now says that Anthropic's cost report leaves out Priority Tier spend, so an organisation on that tier really spends more than the figures show

0.12.1 (2026-09-07)

• Fixed: The last-update stamp of an account no longer moves forward while the provider is only throttling — it dates the values standing next to it, so you can see how old they really are
• Improved: Twenty-five more datapoints explain themselves in the object tree — what "today" means (the provider counts it in UTC), and why the cost totals can be lower than the accounts show

0.12.0 (2026-09-06)

• Fixed: An account that has not been signed in yet no longer reports a rejected sign-in — no warning, no notification, and the settings page keeps offering the sign-in button
• Fixed: An account whose API key is missing or unreadable is now shown as not delivering, instead of leaving its old values standing as though they were current
• Fixed: An answer arriving while the adapter shuts down can no longer mark an account as online again after the shutdown wrote it offline
• Fixed: A throttled account counts as delivering everywhere now — the connection icon and the "reachable accounts" total no longer contradict each other
• Fixed: A limit the provider reports as empty is no longer shown as 0 % used, and a Google quota without a value no longer reads as completely used up
• Fixed: A rejected ChatGPT sign-in now says so at once instead of leaving you waiting for a quarter of an hour, and a Google account keeps delivering when one route is unavailable
• Fixed: A Google account without an AI subscription says so, instead of asking for a sign-in that cannot change the answer
• New: Every limit window shows whether it is the limit currently in force — with Claude the provider states it, elsewhere it is the window that speaks for the account
• Improved: An account is reported as at its limit when the provider says the window is closed, not only when the percentage happens to reach 100
• Improved: A window's reset time is written to the minute, so a recording of it no longer gains an entry on every single query, only on real changes
• Improved: An account that is delivering again says so in the log, instead of leaving the warning about its outage standing as the last word on it
• Improved: The settings page no longer asks the adapter for every status every four seconds — the values now arrive on their own as they change
• Changed: "Balance sufficient for calls" now sits under credits, where it belongs; the datapoint at the old place is removed automatically
• Changed: Each account node shows the readable provider name instead of the internal one — "Claude Max (Claude)" instead of "Claude Max (claude-sub)"
• Fixed: A per-model folder is now named in your ioBroker language as well, instead of carrying the provider's bare model identifier as its only name
• New: The datapoints whose meaning is not obvious from their name now carry a short explanation in eleven languages, shown in the object tree

0.11.0 (2026-09-05)

• Fixed: Signing in from the instance settings works again — a leftover setting from an earlier version had silently closed the adapter's message channel, so none of the three flows reached it
• Fixed: A subscription whose stored sign-in was rejected no longer claims to be signed in — the row now offers the sign-in again instead of showing a green check next to an error
• Fixed: The status badge of an account no longer blanks out for a moment when a single status read is missed — a hiccup in the settings page is not an account without a status
• Fixed: A stored credential whose name sorts high in the alphabet is no longer missing from the account list in the instance settings
• Fixed: The settings page falls back to English for a browser language the adapter does not ship, instead of passing that language on unchecked
• Improved: All object names are now available in eleven languages instead of English only, and a renamed object reaches installations that already exist
• Improved: ChatGPT usage is read with the identity that endpoint expects, the way the Claude query already did — fewer rejected requests on that account
• Improved: Monthly cost reports can no longer be cut short in silence — a report that does not fit is reported in the log instead of producing a figure that is too low
• Changed: "Highest account utilisation" says what it always measured — the fullest limit window or the account's remaining budget

0.10.0 (2026-09-01)

• Fixed: The reset-time datapoint of a limit window no longer disappears and reappears — it stays and simply empties while no window is running
• Fixed: The settings page no longer shows the sign-in screen to a signed-in account, and its rows load without waiting for the credential storage scan
• Improved: Claude usage is read with far fewer rejections — the query now identifies itself the way the endpoint expects
• Changed: New Claude sign-ins request only the profile permission — the stored access can no longer create API keys or run models
• New: ChatGPT accounts show their purchasable limit-reset credits — how many are available and when the next one expires
• Improved: An unreadable provider answer is now reported as a service fault instead of a missing connection

License

MIT License

Copyright (c) 2026 krobi krobi@power-dreams.com

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Developed with assistance from Claude.ai