API keys

POST /keys

{
  "project_id": "<uuid>",
  "name": "checkout-summariser-prod",
  "expires_at": "2026-12-31T00:00:00Z",
  "credit_limit_cents": 5000000000,
  "allowed_models": ["IntelligentInference/gpt-oss-120b(Global)"],
  "environment": "production"
}

project_id and name are required; the rest are optional. project_id must be a project in the caller's organisation (404 otherwise, so a session for one organisation cannot mint keys against another's project).

Response:

{
  "id": "<uuid>",
  "name": "checkout-summariser-prod",
  "project_id": "<uuid>",
  "key_prefix": "sk_i2_XXXXXXXX",
  "rate_limit_rpm": 60,
  "expires_at": "2026-12-31T00:00:00+00:00",
  "credit_limit_cents": 5000000000,
  "allowed_models": ["IntelligentInference/gpt-oss-120b(Global)"],
  "environment": "production",
  "enforcement_status": {
    "expires_at": "enforced",
    "credit_limit_cents": "enforced",
    "allowed_models": "enforced"
  },
  "api_key": "sk_i2_<32 characters>"
}

Important api_key is the secret and this response is the only time it exists in the clear. It is stored hashed; GET /keys returns key_prefix only.

rate_limit_rpm is not an input: every key is issued at the platform default of 60 and the field is echoed so you can see what applies. enforcement_status states whether the gateway currently enforces each limit; all three are enforced on the request path today, and the field exists so a client never has to assume that from the names alone.

GET /keys

{
  "keys": [
    {
      "id": "<uuid>",
      "name": "checkout-summariser-prod",
      "key_prefix": "sk_i2_XXXXXXXX",
      "project_id": "<uuid>",
      "created_at": "2026-08-04T10:38:49.451247+00:00",
      "last_used_at": null,
      "rate_limit_rpm": 60,
      "expires_at": null,
      "credit_limit_cents": null,
      "allowed_models": null,
      "environment": "production"
    }
  ]
}

Newest first, the caller's organisation only, never a hash or a secret. last_used_at is derived at read time from the request log rather than stored on the key, so null means the key has genuinely never been used. If the analytics store is unreachable when you call this, every row's last_used_at degrades to null for that one response instead of failing the call: retry before concluding a key is unused.

DELETE /keys/{key_id}

No body. Revokes the key; the gateway refuses it on its very next request. 404 for a key that is not the caller's, does not exist, or is already revoked (never a silent 200). Writes an api_key.revoked event to the activity log with the key's prefix as the target. Returns {"message": "API key revoked"}.

PATCH /keys/{key_id}

Changes a key's rate limit, credit limit and expiry. Name, project, model scope and environment are still fixed at creation; to change any of those, mint a new key and revoke the old one.

{
  "rate_limit_rpm": 1000,
  "credit_limit_pkr": 500,
  "expires_at": 1790333428
}

This is a full replace, not a sparse patch. All three fields are evaluated on every call, so send the values you want the key to end up with, not just the ones you are changing.

Two of them deliberately differ from the same concepts on POST /keys, which is worth reading twice before you wire a form to this:

Two invariants are enforced server-side, both 400, and each error names the real figure so a form can show it without a second call:

1. Spend floor. A key's limit cannot be set below what it has already spent this cycle.
2. Headroom ceiling. This key's new limit plus every other active key's limit cannot exceed the organisation's budget. A key with no limit of its own contributes nothing to that sum.

Every key in the organisation is locked for the duration of the call, so two edits to different keys fired at the same moment are serialised rather than each reading the other's stale total. You do not need to queue or debounce them client-side.

Returns 200 with the stored values, credit_limit_cents back in micro-PKR:

{
  "id": "<uuid>",
  "key_prefix": "sk_i2_XXXXXXXX",
  "rate_limit_rpm": 1000,
  "expires_at": "2026-09-25T00:00:00+00:00",
  "credit_limit_cents": 500000000
}

404 for a key that is not the caller's. Writes an api_key.updated event to the activity log.