API reference

Everything lives on one base URL:

https://api.proxyllm.ai

Authenticate with the Authorization header:

Authorization: <token>

The Bearer prefix is optional; both forms work. Two kinds of token exist: a routing key (pllm_…) for inference and an account token (sk_… read+write, rk_… read-only) for management. Authentication covers both, including read-only scopes and what happens when you mix them up.

Inference

OpenAI-compatible, authed by a routing key. Point any OpenAI SDK at https://api.proxyllm.ai/v1.

• POST /v1/chat/completions - Chat Completions, including streaming and tool calls.
• POST /v1/responses - Responses API.
• GET /v1/models - the model catalog, including the flagship and mini tiers.
• GET /v1/key - introspect the routing key: lanes, budget, Codex status.
• POST /blitz - run up to 50 prompts through the key’s routing in one request.

Account management

Account endpoints live under /v1/organizations/, the same shape OpenAI and Anthropic use for their admin APIs. Authed by an account token.

• GET /v1/organizations/me - who this token belongs to and its scope.
• GET /v1/organizations/usage - 30-day usage summary.
• GET | POST /v1/organizations/routing-keys - list or create routing keys.
• GET | POST /v1/organizations/bridges - list or register Claude Code bridges.
• GET | POST | DELETE /v1/organizations/codex-session - Codex connection status, connect, disconnect.

The older /v1/routing-keys, /v1/bridges, and /api/codex?action=… paths still work as legacy aliases; new integrations should use /v1/organizations/….

Error shapes

Inference endpoints return OpenAI-shaped errors:

{
  "error": {
    "message": "Routing key monthly budget exhausted",
    "type": "budget_exceeded",
    "code": "budget_exceeded"
  }
}

Management endpoints and POST /blitz return a flat shape:

{
  "error": "This account token is read-only. Use a read and write token (sk_...) from Settings for this action.",
  "code": "read_only_token"
}

Messages are plain sentences that say what to do next. A few validation errors (missing model, messages, or input on inference) return the OpenAI shape with no code.

Error codes

Every code the gateway returns, across all documented endpoints:

Status	Code	When
400	bad_request	A management request body failed validation. The message names the field.
400	no_providers	The routing key has no lanes configured. Add lanes in the dashboard.
400	no_credential	A streaming request landed on a lane whose provider has no saved API key.
400	unsupported_feature	A Responses API parameter the gateway does not support (previous_response_id).
401	signup_required	No key was sent on an inference endpoint, or the key is a raw provider key (sk-…) where a routing key belongs.
401	invalid_api_key	The routing key does not exist or was revoked. Also returned, with an explanatory message, when an account token was pasted where a routing key belongs.
401	invalid_account_token	The account token does not exist or was revoked.
401	auth_required	No token was sent to /blitz.
401	(none)	No token in the Authorization header, on a management endpoint.
402	budget_exceeded	The routing key’s monthly budget is exhausted.
403	read_only_token	A read-only (rk_) token called a write endpoint.
405	method_not_allowed	Wrong HTTP method on a management endpoint.
405	(none)	Wrong HTTP method on an inference endpoint or /blitz.
500	env_missing	The gateway is misconfigured. Not your fault; contact support.
500	server_error	A database operation failed. Details stay server-side.
500	encrypt_failed	Encrypting an uploaded Codex sign-in failed.
500	unhandled	Unexpected crash. Contact support if it repeats.
502	upstream_error	Every lane failed, or the upstream provider broke mid-request. The message carries the last provider error.