fit · before you integrate
When a tab beats a payment
A tab is not a better payment. It is a different instrument, and it is wrong for most things. Five tests decide it — a use case has to pass all of them.
The fit test
- Same counterparty, repeatedly. Netting is pairwise. One payer and one payee, many exchanges — not thousands of distinct sellers, which would be thousands of settlements.
- The unit price is under a few cents. Above that, paying per call is simpler and the batching saves nothing worth the delay.
- Software can measure the unit. Tokens, calls, seconds, bytes, requests. If a human has to adjudicate what was used, it cannot be metered.
- Hours of settlement delay are fine. Defaults are 1.00 USDC or 12 hours. If the counterparty must be paid before the next call, this is the wrong tool.
- Both sides are software. No checkout, no card, no invoice, no accounts-payable department.
Fail two or more and you want a plain transfer, a payment channel with a deposit, or ordinary invoicing — not this. Saying so is cheaper than a bad integration.
Eight that pass
1. LLM inference
An agent making hundreds of completions a day against one provider. Each call is fractions of a cent; a transaction per call costs more than the call.
Why a tab: 1,000 completions at $0.001 is $1.00 — one settlement, one fee, one gas payment, instead of 1,000 of each.
2. API and data queries
Prices, weather, geocoding, enrichment, search. The classic x402 shape: an endpoint worth a tenth of a cent per call.
Why a tab: the seller stops running settlement infrastructure per request, and the buyer stops signing a transaction per request.
3. Agent subcontracting
An orchestrator hires specialists — research, scraping, code execution, verification. Each subagent is a counterparty with a running relationship.
Why a tab: the orchestrator settles once per subagent per cycle instead of per delegation, and the audit trail lives in a ledger it already owns.
4. Compute and GPU rental
Inference endpoints, sandboxes, batch jobs. Usage is continuous and verifiable from the provider's own side.
Why a tab: billing granularity is finer than any on-chain transfer could afford to be, so the ledger carries the precision and the chain carries the settlement.
5. Data feeds and sensor streams
Oracles, telemetry, market data. High frequency, tiny value, one counterparty for months.
Why a tab: the age threshold matters more than the amount here — a feed might never reach 1.00 USDC quickly, but it still needs periodic settlement.
6. Pay-per-crawl and retrieval
Crawlers paying publishers, RAG pipelines paying for retrieval, datasets paid per document served.
Why a tab: a crawler hitting one publisher 10,000 times a day is the purest form of this problem — and the pairwise shape means the publisher never has to trust a platform middleman.
7. Enrichment microservices
KYC, sanctions screening, translation, transcription, address validation. Called in bursts by a service that has one vendor for each.
Why a tab: vendor sprawl — five vendors, five tabs, five settlements per cycle instead of five integrations and five invoices.
8. Model routing and marketplaces
A router spreads work across providers and owes each of them a fraction of every request it forwards.
Why a tab: the router's own revenue is collected per request, so paying providers per settlement is the only structure with positive margin.
When it is the wrong tool
| Situation | Use instead |
|---|---|
| One-off purchase, any size | A plain transfer — the ledger adds nothing but delay |
| Escrow, or a payer you cannot trust at all | A payment channel with a deposit, or escrow (ADR-0002) |
| Each counterparty transacts once | Per-call payment; pairwise netting never compounds |
| Must clear before the next call | A channel, or a card on file — settlement is deliberately slow here |
| A human decides what was used | Invoicing. Unmeterable usage is unbillable usage |
| Amounts in the thousands | Direct transfer; 50 bps is real money at that size |
The arithmetic
Take an agent making 1,000 calls at $0.001 to one provider. Per-call settlement is 1,000 transactions; a tab is one settlement with two transfers. The fee is the same 0.5% either way — what collapses is the number of times you pay gas, wait for confirmation, and poll for a receipt.
The threshold is the dial. Raise it and settlements get rarer and cheaper but the payee's exposure grows; lower it and you are back to paying per call. 1.00 USDC and 12 hours are starting points, not doctrine.
Next
How it works for the flow and the money path · Quickstart to integrate · protocol.json for the machine facts.