Variables
The checklist reports
default_llm as configured only when both the model
id and the key are set — a key alone points at nothing.
The cap
The shared model is a key you pay for, and/api/config announces it to
anonymous callers. So the open-source build caps it by default rather than
leaving that to whoever installs a brick: one million tokens per user and
twenty million for the deployment, over a rolling 24-hour window.
Only consumption billed to the shared key counts. A user who brings their own
API key pays for it themselves and is never capped by these settings.
The window rolls; it does not reset on the hour or at midnight. A calendar
reset would let one account spend a full allowance at 23:59 and another at
00:01 — the ceiling that matters is what can be spent in any 24 hours, not what
a counter reads.
The second cap exists because the first one is only worth as much as the number
of accounts is bounded. Self-registration is open by default
(AUTH_REGISTER_ENABLED), so a per-user cap is one signup away from being
lifted. The deployment cap is not.
A refused run answers 402 before anything starts, with a body an interface
can use:
scope is user or deployment, and retry_after_seconds is the wait until
the oldest counted consumption leaves the window — the first moment budget
comes back.
Set a cap to 0 to remove it. If a commercial brick supplies its own ceiling,
it takes over entirely and these settings are ignored: the core steps aside for
the dedicated extension point, not for any add-on that happens to register a
run guard.
Should the consumption table be unreadable, the run is allowed and a
warning is logged. A hiccup in accounting must not make the product mute — but
the opening has to be audible.
What users see
The counter under the sidebar shows the tokens spent this month on the shared model. The bar and percentage come from a commercial brick, which supplies a monthly ceiling per user or per plan; without it the sidebar shows a counter only. That gauge and the cap above are separate: the cap protects the bill, the gauge tells a customer where they stand in their plan.Check
default_llm_available: true and a default_llm_model_id mean agents can run
with no key of their own. This endpoint is public and deliberately says nothing
about the key itself.