Every control that used to hide in a vendor portal.

Budgets, rate limits, key hygiene, model routing, and an immutable ledger — one system, not a stack of screenshots in chat.

SPEND POLICY

Hard caps, soft caps, and a graceful step-down.

Workspace monthly. Team weekly. Per-key daily. Per-model request and token rates. Near the line we can warn, throttle, switch to a smaller model, or refuse with an error your app already understands.

TEAM · SEARCH91%

TEAM · SUPPORT38%

KEY · contractor-08blocked

VIRTUAL KEYS

Give the app a credential. Keep the real one in the vault.

Mint scoped keys with model allowlists, origin rules, and IP ranges. Rotate upstream secrets without touching application config. Contract access can expire on Friday without a war room.

vk_live_web frontier, balanced
vk_eval_nightly small, flash
vk_agency_q3 expires 21d
sk_upstream_prod vaulted

ROUTING

A written policy chooses the model, not a hope.

Describe the lane in plain language: if the prompt is long, use long-context; if the primary lane returns 429, fall to fast-lane; never send private text to an unmanaged fine-tune. Shadow a candidate before you cut over.

route support.prod

primary → balanced-text
overflow → frontier-small
outage → fast-lane
private_mode → regional-eu

OBSERVABILITY

Cost is a first-class metric.

Token burn, cache hits, retry storms, and the estimated invoice — by product, environment, and person. Stream it out, or keep it here. Either way, finance and engineering finally share a number.

Walk the ledger with realistic numbers.

The preview workspace is loaded. No account required.

Open dashboard