v0.8.5 · Beta · Celo mainnet

Every cap a buyer can hit.

Rate limits, body sizes, deadlines and ceilings, one number each, taken from the code that enforces them. A limit is refused before payment unless a row says otherwise. The safety notes that go with these numbers are in Limits and safety.

Free SQD route

gateway.usebuy.ai/sqd/* proxies the SQD Portal with no 402. It is the only route without a price, so it has the tightest request limits.

Limit	Value	When hit
Requests per IP	120 per minute, burst 20	`429` with a `Retry-After` header
Methods	`GET` and `POST`	`405`
POST body	64 KiB	`413`
Upstream deadline	30 s	`504 gateway_timeout`; nothing was paid

Every /sqd response carries RateLimit-Limit (the burst, 20), RateLimit-Policy (120;w=60), RateLimit-Remaining and RateLimit-Reset (seconds until the burst is back); a 429 adds Retry-After. An error from the SQD portal comes back with its status in the gateway envelope (upstream_error, a bounded excerpt, a correlation id).

Paid capability routes

gateway.usebuy.ai/v1/*: the capabilities listed at /v1/catalog. Each request is quoted with a 402 and paid with one X-PAYMENT header.

Limit	Value	When hit
JSON request body	64 KiB	`413` before the 402
Request headers, `X-PAYMENT` included	16 KiB in total	`431` from the server before any route runs
Payment validity (`maxTimeoutSeconds` on the 402)	300 s	sign and send within it
Upstream deadline	150 s	`504 timeout`, `retryable: false`; do not retry automatically
Whole request	12 min	the connection is closed

Prices are in US dollars, payable in USDC, USD₮ or USA₮. The 402 challenge for the exact request is authoritative; the catalog only predicts it.

Partner routes

Resold on the same host. Each route has two rate limits: one per caller IP and one for the whole route, which protects the partner. Both answer 429 with Retry-After before payment. A 429 the partner itself returns comes back after settlement, so that request was paid for.

Limit	`/chainstack/celo`	`/cencori/v1/chat/completions`
Price per request	0.01 USDC	0.01 USDC
Per IP	300 per minute, burst 50	120 per minute, burst 20
Whole route	60 per minute, burst 10	60 per minute, burst 10
Methods	`POST`	`POST`
Request body	128 KiB, valid JSON	32 KiB, a JSON object
Per body	up to 2000 JSON-RPC calls in one batch	one completion; `model` must be `maximo-atlas-1.3`, `max_tokens` must be 256, 1024 or 4096
Fields	any JSON-RPC	only `model`, `messages`, `max_tokens`, `temperature`, `top_p`, `stop`, `seed`
Upstream deadline	30 s	60 s

A body that breaks a rule above is refused with a 400 before the 402.

Virtual machines

usebuy.ai/google/vm and usebuy.ai/google/ssh. The fleet caps are shared by everyone; a request over the VM count is refused with 503 at_capacity and one over the hourly spend with 503 at_spend_capacity, both before payment.

Limit	Value
VMs at once, fleet-wide	100
Catalogued compute, fleet-wide	$25 per hour; admission stops above it
Lease per purchase or renewal	1 hour
Total life of one VM, renewals included	24 hours
Network egress per VM	10 GB; above it the VM is deleted within minutes, the lease is not refunded, and the poll URL reports `egressCapExceeded: true`
Request body	64 KiB
VM type	on-demand only

Sizes, prices and which sizes need a Self proof are in Rent a VM.

Client-side guards

These live in the buy CLI and MCP server, not on the gateway. They stop a request before anything is sent.

Guard	Value
`--max-amount`	A decimal token amount, e.g. `0.05`; a quote above it is refused unpaid. Without the flag the CLI pays what the 402 quotes, so set it close to the quote.
Implausible `--max-amount`	1000 or more is refused without `--force`: the 402 quotes atomic units (`16753` is 0.016753 USDC), so a copied quote reads as 16,753 USDC.
`buy account cap <name> <amount>`	A daily spend cap per account in stablecoins. A cap of 1000 or more needs `--force`.
MCP `maxAmount`	Atomic six-decimal units, the same figure as the 402 quote.

$ buy account cap demo 0.50