Compute chat
/app/compute talks to a model from inside the desk. Usage spends compute credit. You never paste a vendor key.
When to use chat
Use chat when you have already converted a marker onto the compute rail and you want an answer without wiring up an SDK. Pick a vendor and a model on the composer and send. The desk meters the spend against your remaining cents and stops when they run out.
All 10 vendors on the control are the same rail. They are not separate balances — switching from Claude to Gemini does not reset anything.
Threads
- Threads are stored in this browser, not on the desk.
- Clearing site data clears your threads. The spend is already recorded.
- The desk does not train on your messages and does not retain them beyond the request needed to answer.
How metering works
Every request is priced on the vendor’s published token rates and rounded up to the nearest cent. A one-line prompt still spends at least one cent. The remaining balance is shown live in the composer, and a request that would exceed it is refused before it is sent upstream rather than failing halfway through a stream.
Limits
| Limit | Value |
|---|---|
| Minimum charge | 1 cent per request |
| Context | Whatever the locked model supports |
| Streaming | On by default, cancellable mid-stream |
| Rate limit | 20 requests per minute per address |