Endpoints

Every endpoint Faucet serves, which credential authenticates it, and what it answers with.

Everything below is reachable at https://api.intfaucet.com. Two endpoints are public; the rest take a key in Authorization: Bearer.

Inference

Endpoint	Notes
`POST /v1/chat/completions`	Streaming and non-streaming, every chat provider. See Chat
`POST /v1/messages`	Anthropic's native format. `anthropic/*` chat models only
`POST /v1/messages/count_tokens`	Anthropic's native token count. Admitted and metered at zero — the one deliberately free operation
`POST /v1/embeddings`	Never streamed. See Embeddings
`POST /v1/images/generations`	Never streamed. See Images
`POST /v1/images/edits`	JSON or `multipart/form-data`
`POST /v1/audio/transcriptions`	`multipart/form-data` only. See Transcription

A model answers on one endpoint. Sending it to another is a 400 naming the endpoint that would have worked — not a 404, because the slug is fine. See Choosing a model.

Catalog

Endpoint	Notes
`GET /v1/models`	Your catalog, filtered to what this key can reach. Exactly OpenAI-shaped
`GET /v1/models/{id}`	Accepts a slug or an alias and echoes the canonical slug, plus a `faucet` capability block

GET /v1/models is narrower than the Models reference: it applies the key's allowlist and skips providers holding no usable credential, so it never advertises a model that would 403 or 503 on use.

The detail route answers 404 when the model is not carried, is outside your key's allowlist, or has no usable credential — one answer for all three, because distinguishing them would confirm which model ids exist. What the faucet block contains is on Model capabilities.

Money and credentials

Endpoint	Credential	Notes
`GET /v1/credits`	inference key	The organization's live credit. `available` is what the next request is judged against
`GET /v1/key`	inference key	What the presenting key is allowed, and its own spend this month
`DELETE /v1/key`	inference key	Revoke the presenting key. Names no id; idempotent
`POST /v1/management/keys`	management key	Mint an inference key
`GET /v1/management/keys`	management key	This organization's keys
`GET /v1/management/keys/{id}`	management key	Its limits, plus `spend_this_month_micros`
`PATCH /v1/management/keys/{id}`	management key	Change a plan
`DELETE /v1/management/keys/{id}`	management key	Revoke

Neither credential works where the other does

A management key (fct_mgmt_…) is refused at every inference endpoint, and an inference key is refused on /v1/management/*. A key that leaks from a subscriber's machine can spend its budget and cannot mint more. See Management keys.

/v1/credits and /v1/key are not rate-limited or metered. A key with nothing left must still be able to ask what it is allowed — admitting there would refuse the one request whose purpose is explaining a refusal.

The two routes answer different questions, and the difference matters under a reseller: /v1/credits reports the organization's balance, which under a reseller holds every subscriber and is the reseller's commercial information. /v1/key reports only the presenting key's own budget and spend, which is what a subscriber may see. Both are scoped to the key that calls them, so the worst a stolen key learns is the allowance of the credential the thief already has.

Public, unauthenticated

Endpoint	Notes
`GET /health`	That this process is alive. A constant; touches no database
`GET /status`	Whether the gateway is working, per provider

These are not interchangeable. /health is what liveness probes watch — a liveness check that read the database would turn a database blip into a restart loop. /status reads recent traffic, so it can be slow or fail without anything being restarted.

/status is the only endpoint that answers cross-origin, and deliberately: its payload is public, identical for every caller, and sent with no credentials. Nothing under /v1 is callable from a browser, because a Faucet key is a server-side credential.

{
  "status": "ok",                  // or "degraded"
  "windowMinutes": 30,
  "providers": [
    {
      "slug": "anthropic",
      "status": "ok",              // ok · degraded · unknown
      "sample": "normal",          // none · low · normal
      "reportedStatus": "operational",
      "statusPage": "https://status.anthropic.com"
    }
  ]
}

Field	Meaning
`status`	`ok` or `degraded`, across every provider
`windowMinutes`	How far back the observed verdict looks
`providers[].status`	`ok`, `degraded` or `unknown` for one provider
`providers[].sample`	`none`, `low` or `normal` — how much traffic the verdict rests on
`providers[].reportedStatus`	The provider's own published report, or `null`
`providers[].statusPage`	Where to read that report yourself, or `null`

Three things about that payload:

It never reports request counts or model slugs. sample buckets how much traffic a verdict rests on without saying how much; low is what stops you trusting a degraded off a handful of requests.
Observed health wins. A provider's own report changes the aggregate only when the observed sample is empty. A failed or unreadable provider feed is null, never a degradation.
unknown alone is not degraded. The top-level status is degraded when a provider is observed degraded, or is unknown and independently reports a problem.

Errors

Errors follow the dialect of the endpoint: Anthropic's envelope on /v1/messages and its token-counting route, OpenAI's { "error": { "message", "type", "param", "code" } } everywhere else — transcription included. Status meanings, the three 402 scopes and your allowance are on Errors and limits.

Inference

Catalog

Money and credentials

Public, unauthenticated

Errors

On this page