Credits and wallets
One credit is one cent. Spend comes out of two separate wallets, both held at the organization level, and every paid response tells you what it cost and what is left. Failed calls cost nothing, so a retry loop cannot quietly drain a balance.
Two wallets
Data spend and model spend are metered separately, because they scale for different reasons. Both wallets belong to the organization and are shared by every workspace inside it, so moving work between workspaces never strands a balance.
| Wallet | Pays for | Overdraft |
|---|---|---|
| data | Curated data sources and every other per-call data charge, priced per call, per endpoint. Your-credential connections do not draw on it: they are recorded at zero. | none, the call is refused first |
| ai | Model input and output tokens, cache reads and writes, tool use, and managed-agent runtime seconds. | may dip slightly negative |
Per-call prices for the curated sources are not fixed platform-wide. Each endpoint carries its own cost: read it from GET /v1/platforms or on the endpoint page in the API reference. The workspace database is not metered at all: the /v1/db/* routes deduct no data credits.
How a charge is applied
A data charge is resolved in three moves, in this order.
- 01The organization discount is applied first. It is the best discount across every active subscription, so a credit package covers every data call the organization makes. Builder plans carry no discount: their price already includes the data credits.
- 02The monthly subscription allowance drains next. Credit-package allowances go first, oldest first, and the plan's included credits last. Allowances refill each billing period and do not carry over.
- 03Whatever is left debits the purchased balance, which never expires. The trial credits land here.
The discount applies to data spend only. AI spend drains the plan's AI allotment first and then the purchased AI balance, at the metered rate. Builder plans include no AI allotment, so their AI spend comes straight from what you top up.
the two overdraft rules differ
The data wallet cannot go negative, and a data charge is all or nothing. Before contacting the upstream provider, a call holds what it may cost against your allowance plus purchased balance, minus what calls still running hold. If that cannot cover it, the call is refused and costs nothing, so a burst of parallel calls cannot all spend the same credits. A call that finishes is charged what it actually cost and its hold is released.
When your balance alone covers the call and only your other running calls' holds are in the way, the refusal is RATE_LIMITED (429) with retryAfter: nothing is spent, and the call goes through once those holds settle.
When the balance itself is short, on a builder plan or the trial the refusal is DATA_CREDITS_EXHAUSTED (402) carrying an offer for the next plan. Everywhere else it is INSUFFICIENT_CREDITS (402) with the required and available amounts.
The AI wallet is allowed to dip slightly negative. A response already streaming is never cut mid-token to settle a rounding race; the small shortfall lands on the purchased AI balance and clears on the next top-up.
Failed calls cost nothing. A data call that errors is recorded at zero cost, and calls through your own vendor credentials are always recorded at zero because you are paying the vendor directly.
Starting credits
A new account starts on the no-card trial: its first verified sign-in puts $3 on the data wallet and $1 on the AI wallet, once, alongside Growth limits. The AI credit is for hosted agents. The data credit is enough to make real calls and see the metering work end to end before any card exists. How the trial runs, and when it ends, is on Plans and limits.
Bounties add more. Each is an action that helps superagnt grow, like following on X or starring one of our GitHub repos, and each pays data credits once per organization, and once per account across all organizations. Claim them on the Bounties page in the dashboard, in the onboarding flow, or have your agent claim them for you with agnt_bounties_list and agnt_bounties_claim. Bounties are on the honor system: nothing is checked, and each one pays when you claim it.
| Bounty | Pays |
|---|---|
| Follow on X | $0.25 |
| Star superagnt/plugins on GitHub | $0.25 |
| Star superagnt/leverage on GitHub | $0.25 |
| Follow on LinkedIn | $0.25 |
| Subscribe on YouTube | $0.25 |
| Tell us where you heard about us | $0.25 |
A credit package is a monthly data-credit allowance that also carries a standing discount and a one-time signup bonus. The bonus lands on the purchased balance and never expires. Packages sit on top of a managed plan, or stand alone for an organization with no plan; an organization on a builder plan or the trial cannot add one.
| Package | Price / mo | Data credits / mo | Discount | One-time bonus |
|---|---|---|---|---|
| Data Credits Small | $29 | $29 | 10% | $10 |
| Data Credits Standard | $99 | $99 | 15% | $33 |
| Data Credits Large | $299 | $299 | 20% | $100 |
Buying credits
Top-ups happen in the dashboard at app.superagnt.com. You pick the wallet, then the amount. The minimum purchase is $10, and the quick-buy presets are $25, $50, $100 and $250; any amount at or above the minimum works.
builder plans buy ai only
- Purchased credits never expire. They sit on the organization balance until something spends them.
- Top-ups are per wallet. Buying data credits does not fund model spend, and the reverse is also true.
- Invoices, receipts and saved cards live in the Stripe billing portal, opened from the billing page.
Auto-reload
Auto-reload is configured per organization and per wallet: a threshold and a reload amount, each wallet independently. When a debit takes the balance below the threshold, the saved card is charged for the reload amount. It needs a card already on file, and the reload amount is subject to the same $10 minimum as a manual purchase. On a builder plan or the trial only the AI wallet reloads; the data wallet never does.
A short cooldown sits between consecutive attempts on the same wallet, currently 60 seconds. That is what stops a burst of concurrent calls from queueing several charges at once: the first debit to cross the threshold claims the slot, and its siblings do nothing.
A declined card is not retried every minute. After one decline the wallet waits an hour before the next attempt, and after two, six hours. A third decline in a row, or a charge the bank wants to confirm (which can’t happen while auto-reload runs in the background), pauses auto-reload on that wallet: the card isn’t charged again, every workspace gets a notice, and the organization owner gets one email. Updating the auto-reload card on the billing page turns it back on.
Who can spend
Buying credits, changing auto-reload, changing the plan and adding seats are restricted to organization owners and admins. A member who tries gets a 403 saying so. Everyone in the workspace can read the balance and the usage history.
agents never move money
Reading your balance
GET /v1/credits returns the data wallet: the purchased balance, the remaining subscription allowance, and the recent credit-transaction ledger.
curl "https://api.superagnt.com/v1/credits" \
-H "Authorization: Bearer $AGNTDATA_API_KEY"You rarely need to poll it, because every paid call already carries the balance back inline. A successful response puts the cost of that call and both data-wallet figures in meta.
{
"success": true,
"data": { /* ... */ },
"meta": {
"costCents": 1.5,
"purchasedBalanceCents": 4248.5,
"subscriptionRemainingCents": 1800,
"cached": false,
"latencyMs": 412
}
}From inside an agent session, agnt_credits_balance returns the same data-wallet figures plus their total, so an agent can check whether it can afford a batch before starting one. During the trial it also reports where the trial stands. For AI spend and a full breakdown across both wallets, see Usage.