Connect
Enter the gateway master key. It is kept in this page's memory only and sent as X-Admin-Key on each request — store it in your password manager, not here. The gateway host is a label for that vault entry (so test and prod stay separate); it is never sent to the server.
Job history
Every orchestrated job, with the tokens it spent and what that cost. Cost is joined from the price table at read time — change a price and every past row re-costs itself.
| Submitted | Service | Identity | Outcome | Delivery | Duration | Tokens | Cost |
|---|
No jobs match these filters.
Dead letters 0
Results whose delivery failed. Replay re-delivers the result that already exists — it does not re-run the job, so no tokens are spent. Entries that never parsed can only be discarded, and while one sits here the DLQ alert stays raised.
| Entry | Job | Service | Reason |
|---|
Empty — nothing has been dead-lettered.
Chargeback summary
Tokens and cost per system / tenant / model over the window. Cached is part of Input (and reasoning tokens are part of Output): the columns overlap on purpose, and each token is billed once at its own rate.
| System | Tenant | Model | Calls | Input | Output | Cached | Total | Cost |
|---|
No usage recorded for these filters.
Model prices
Prices are pulled from Langfuse, which maintains model definitions and updates them with its releases — nobody edits a price list here. The sync runs on its own timer and writes rows dated by Langfuse's own start date. A row publishes from its date, so for each model and kind the newest row bills and older ones are shadowed. Langfuse often dates a definition from the model's launch, which means a manual row keeps overriding the synced price indefinitely — retract it to hand that model back to the sync.
| Model | Kind | Price / token | Per 1M tokens | Currency | Effective from | Source |
|---|
No prices yet — the sync has not run, or Langfuse knows none of the models in use.
Override a price
For a contracted rate or a currency conversion. An override is simply a newer-dated row: reports pick the newest price not later than the usage day, and the Langfuse sync never writes a row dated later than Langfuse's own start date, so it cannot silently revert this. Prices are per single token — providers quote per million, so divide by 1 000 000 first.
Issue a new key
Issued keys
| System | Permissions | Created | Last used | Expires | Notes |
|---|
No keys issued yet.
Using a key
Call the gateway with the key in the X-Api-Key header (the consumer system is derived from the key itself); X-Tenant and X-User are free identifiers used for usage tracking. Example against the pdf_anonymization service: