Credits and usage

Read the usage page#

The dashboard usage area combines billing context with the resources that can stop an operation: current plan, credit balance, site allowance, live-voice minutes, clone entitlement, top-up choices, and billing actions. Check it before a launch and after adding a high-volume embed, connector, or API client.

Do not interpret one number as a universal quota. Each entitlement protects a different resource.

Distinguish balances and limits#

Credits do not buy additional API throughput, sites, live minutes, or clones. An entitlement does not guarantee the credit/provider path required to complete the operation.

What consumes credits#

Metered work can include site/agent generation, normal agent answers, knowledge embeddings, media transcription, text-to-speech, connector voice, and voice cloning. Cost varies with provider, model, prompt/context length, output length, audio duration, and operation type. One “message” therefore has no fixed universal credit price.

Owner preview/test actions can consume real resources when they call live providers. Use short, representative acceptance tests and avoid repeatedly indexing the same large source or cloning the same sample while debugging an unrelated issue.

Read usage history#

Usage history shows the final credits actually consumed by each completed operation. Temporary authorization holds, reserve settlements, and refunds remain internal accounting details instead of appearing as separate customer-facing rows. Very small embedding operations display as <0.001 rather than being rounded to zero.

Other balance changes is a separate table for plan-credit grants, top-ups, expiries, and manual adjustments. It does not duplicate normal AI usage. While an operation is still in progress, the spendable balance can temporarily reflect a hold before the final usage row appears or a failed operation is released; refresh after completion when reconciling a just-finished action.

Monthly and purchased credits#

The monthly allowance comes from the active plan. Unused monthly credits roll over for one extra cycle, then expire. One-time packs add 1,000 credits for $8, 5,000 for $35, or 15,000 for $90; purchased credits do not expire while a subscription is active.

Buying a pack does not restore a canceled/expired platform subscription. Resolve access/billing first when the dashboard reports an entitlement problem rather than only a low balance.

Auto top-up#

Auto top-up buys a credit pack for you when the balance falls to a threshold you choose. Turn it on in Settings → Usage → Auto top-up: pick the threshold (between 50 and 5,000 credits) and the pack to buy. It charges the card already saved with the payment provider; add or change that card through the customer portal link beside the form.

The safeguards are deliberate. Only one purchase runs at a time, and while it settles the page says a purchase is in progress. At most three automatic purchases happen per day. Credits appear when the payment settles, not when the charge starts. If the card is declined no credits are added, the page shows the last error, and one email that day explains what to fix; the charge is not retried until the payment method works.

Auto top-up is available only where off-session charging is enabled and a card is on file. When it is unavailable the page says so instead of showing a switch that would do nothing.

Add-ons#

Add-ons add capacity on top of the current plan. They are billed monthly, can be cancelled at any time, and are managed in Settings → Usage → Add-ons.

An add-on is a separate subscription from the plan, so cancelling the add-on does not touch the plan and cancelling the plan does not quietly keep the add-on. A cancelled add-on runs to the end of the period already paid for. Only the billing owner can buy or cancel one.

Live voice accounting#

Operate includes 30 live minutes per month and Scale includes 90. Each call also has an owner-configured automatic duration limit from 30 seconds to 20 minutes. When a live-call token is issued, the wallet is charged up front for the server-authorized maximum duration, bounded by the remaining monthly allowance. The same duration is held against the live-minute allowance so concurrent calls cannot exceed it.

Live media flows directly between the browser and the real-time provider, so the server cannot independently verify a shorter client-reported duration. Unused time is therefore not partially refunded after a token is delivered; an abandoned or unreported session is charged at its authorized limit. A server-side failure before token delivery releases the charge. End calls cleanly to close the session promptly, but forecast credits using the configured call limit rather than the caller's expected talk time.

Voice messages are different: they consume metered transcription/model/speech work rather than the live-minute entitlement. Diagnose the correct bucket when one mode succeeds and the other fails.

Forecast usage#

1. Choose representative short, normal, and long visitor questions.
2. Test text, knowledge-heavy answers, voice messages, and live calls separately.
3. Record credit/minute changes for the acceptance set.
4. Multiply by expected daily traffic and channel mix.
5. Add headroom for retries, crawls, provider failures, internal QA, campaigns, and traffic spikes.
6. Recheck after changing model, provider, source size, response length, or voice behavior.

Use analytics and conversations to find what caused growth. A traffic spike, looping API client, leaked key, copied embed on an unintended origin, connector retry, or repeated indexing job needs a different fix.

Prevent unexpected usage#

• Restrict embeds to exact intended origins.
• Give API keys the smallest scopes and separate keys by service/environment.
• Make mutations idempotent and honor 429 retry headers.
• Verify webhook signatures and deduplicate event IDs.
• Rotate leaked API/provider/connector credentials immediately.
• Keep public agents bounded by clear persona and response instructions.
• Remove stale automation, preview deployments, connectors, and unused keys.
• Alert operational owners before balances or live minutes approach zero.

Investigate a spike#

Record the affected workspace/site, time range, channel, deployment, and recent changes. Compare usage with analytics, conversation timestamps, API request IDs, webhook delivery logs, and connector last-event/error state. Do not send plaintext keys or visitor content to support unless an approved secure process explicitly requires it.

Contain the source first: disable a connector, revoke a key, remove an unintended origin, or stop a client retry loop. Then preserve safe evidence, rotate credentials where needed, and re-enable one path at a time.

Troubleshooting#

See Plans and pricing for the current entitlement table.