For agents and the people who run them

Inferrail HostedOne budget per run, across your whole fleet.

Give each AI run a dollar ceiling in a request header. Every call in that run shares it, including parallel calls from different agents and machines, and a call that would push the run past it is refused before it reaches OpenAI or Anthropic. You bring your own provider key with each request; we never store it.

Experimental · 2,000 governed runs a month free · then $0.001 per run · no subscription

Not open yet. This page describes the terms Inferrail Hosted will launch with. The hosted service isn't accepting workspaces yet; until it is, the open-source gateway does the same per-run enforcement on your own machine.

Price

Unit

A governed run: a distinct work id (X-Inferrail-Attribute-Work-Id) in a calendar month (UTC). Calls inside a run are not charged separately.

Billed when

At least one call in the run is answered by your provider. A run whose calls all fail, or are all refused by its own budget, costs nothing.

Free

2,000 governed runs per workspace per month.

After that

Prepaid credits at $0.001 per run. People pay by card: $10 for 10,000 runs. Agents pay in USDC over x402: $1 for 1,000 runs, with no account.

No surprises

Credits are prepaid. No subscription, no auto top-up, no overage bill. When the allowance and credits run out, calls are refused before reaching your provider, with a link to buy more.

Model spend

Billed by your provider on your own key. Inferrail doesn't resell model access or mark it up.

Self-hosted

The open-source gateway is free and does everything on one machine: pip install inferrail. Hosted is for when many agents and machines must share one budget ledger and you'd rather not run it.

Start

POST /v1/workspaces returns a workspace key, once. No sign-up form.
Point an OpenAI or Anthropic SDK at /v1 with the workspace key as the API key, and send your provider key in X-Provider-Api-Key.
Name the run and set its ceiling: X-Inferrail-Attribute-Work-Id, X-Inferrail-Budget-Usd.
GET /v1/work/<id> shows what the run cost. GET /pricing and /llms.txt publish these terms for agents.

Past your allowance with no credits left, a call gets HTTP 402 allowance_exhausted, listing both ways to buy. Nothing reaches your provider.

What we guarantee

A run's budget is reserved before each provider call, so concurrent calls can't spend the same dollar. The method and raw results are public: per-run budget benchmark.
Your provider key is read from each request and never written to disk, logged, or returned.
Receipts hold costs and identifiers, not prompts or responses.
Each payment adds credits at most once. A card purchase counts only after Stripe confirms it, and a USDC purchase only after it settles.

Experimental. These prices are a first offer and may change; credits you've already bought keep their terms.