generated: '2026-09-07'
method: searched
source: >-
  https://docs.toapis.com/llms.txt (docs corpus),
  https://docs.toapis.com/docs/cn/api-reference/chat/chat.md,
  https://docs.toapis.com/docs/cn/api-reference/webhooks/task-webhooks.md,
  https://docs.toapis.com/docs/cn/api-reference/rate-limits/async-tasks.md,
  https://docs.toapis.com/.well-known/agent-skills/apis/skill.md
notes: >-
  ToAPIs is an OpenAI-compatible AI gateway; its cross-cutting semantics mirror the
  OpenAI API surface it emulates, with an async task layer for image/video work. No
  OpenAPI is published, so everything here is read from documentation.
auth:
  style: Bearer API key (sk- prefix) in the Authorization header on every endpoint
  see: ../authentication/toapis-authentication.yml
base_urls:
  production: https://toapis.com/v1
  mainland_china: https://toapis.cn (documented drop-in alias for all endpoints)
compatibility:
  - OpenAI Chat Completions API (POST /v1/chat/completions)
  - OpenAI Responses API (POST /v1/responses format, for function calling and server-side context)
  - Anthropic Messages API (native format for Claude-series models)
  - OpenAI Models API (GET /v1/models, scoped to the calling key)
idempotency:
  coverage: none
  note: >-
    No Idempotency-Key or request-replay-protection mechanism is documented on any
    write endpoint. The only deduplication in the surface is receiver-side: webhook
    consumers must deduplicate on the stable event id, and async task submissions
    may carry a client_business_id correlation field — neither prevents a duplicate
    submission from being billed.
reversibility:
  grade: none
  note: >-
    Write surfaces are async generation-task submissions and API-token creation. No
    cancel, void, undo or delete operation is documented for a submitted generation
    task, and no reversal window is stated anywhere in the docs; failed tasks are
    refunded by the provider's own billing (generation.failed is only sent after
    final billing or refund completes), but no caller-initiated reversal exists.
pagination:
  style: none documented
  note: List endpoints published (GET /v1/models) return unpaginated collections; batch task-status queries accept up to 100 task IDs per request.
async_pattern:
  style: submit-then-poll-or-webhook
  submit: POST /v1/images/generations and /v1/videos/generations return a task id immediately
  poll: GET /v1/images/generations/{task_id} and /v1/videos/generations/{task_id}; provider guidance is 5-10s intervals with jitter
  webhook: signed final-state events (see ../asyncapi/toapis-webhooks.yml)
  result_expiry: generated image/video URLs expire after 24 hours
streaming:
  style: 'Server-Sent Events on chat completions via stream=true; async image/video tasks never stream'
error_envelope:
  shape: '{"error": {"code", "message"}} (OpenAI-style)'
  see: ../errors/toapis-problem-types.yml
rate_limit_signaling:
  headers: [Retry-After, X-RateLimit-Limit, X-RateLimit-Remaining, X-RateLimit-Reset, X-RateLimit-Category]
  see: ../rate-limits/toapis-rate-limits.yml
request_id_tracing: not documented
versioning:
  style: URL prefix /v1; webhook payloads carry api_version. No versioning or deprecation policy page is published.
uploads:
  style: multipart upload endpoints (POST /v1/uploads/images, /v1/uploads/videos) return URLs for use in generation requests; base64 inline data is deprecated in favor of upload-then-URL