API & Error Codes
Endpoints and authentication
The Xisu AI base URL and every endpoint: OpenAI-compatible chat/completions, responses, embeddings, images and models, plus Anthropic-native messages, all sharing one key and balance. Includes the auth header, streaming behaviour and dropped streams.
Endpoint address
https://api.xisu.ai/v1Endpoints
The gateway speaks two protocols, on one key and one balance.
OpenAI-compatible (what almost everyone uses)
| Endpoint | Purpose |
|---|---|
POST /v1/chat/completions |
Chat completions |
POST /v1/embeddings |
Text embeddings |
GET /v1/models |
The models this key may call |
POST /v1/images/generations |
Image generation (synchronous; no streaming) |
POST /v1/responses |
Responses API (the Codex CLI and the OpenAI SDK's responses API; models are a subset, see The OpenAI Responses endpoint) |
POST /v1/images/generations |
Image generation (synchronous; no streaming) |
POST /v1/responses |
Responses API (the Codex CLI and the OpenAI SDK's responses API; models are a subset, see The OpenAI Responses endpoint) |
Anthropic-native (works with the Anthropic SDK and Claude Code as-is)
| Endpoint | Purpose |
|---|---|
POST /v1/messages |
Chat completions, with neither request nor response translated |
POST /v1/messages/count_tokens |
Count tokens; not billed |
For the differences, and which one to use, see The Anthropic-native endpoint.
Of the remaining OpenAI endpoints only image generation is offered (/v1/images/generations, synchronous, no streaming, and it needs an image model — whether one is currently on sale is answered by GET /v1/models). Image edits and variations, audio, files, fine-tuning, Assistants and the rest are not offered here, and pricing has no "per audio minute" dimension. (Image input is a different matter — it passes through untouched on /v1/messages.)
Authentication
Standard bearer token:
Authorization: Bearer YOUR_API_KEY
Keys are created in the console, and the full key is shown only once, at creation.
If a key has an IP allowlist, calling from an address outside it returns 401 invalid_api_key — the same code as an invalid key. That is deliberate: a dedicated code would tell whoever holds a leaked key "this key is fine, you're just on the wrong network". So when you debug a 401, check the source IP first instead of rushing to recreate the key.
Streaming
Pass "stream": true. You get standard SSE, terminated by data: [DONE]. That is the OpenAI-compatible shape — /v1/messages returns a native Anthropic event stream with no [DONE]; see The Anthropic-native endpoint. That is the OpenAI-compatible shape — /v1/messages returns a native Anthropic event stream with no [DONE]; see The Anthropic-native endpoint.
- Usage for a streamed response arrives in the final chunk. If the connection drops you won't receive it, but the tokens already generated are still billed — the gateway meters what it produced, not what reached you.
- A long silence usually means the upstream is queueing, not that the connection died. Handle it with your client's own timeout.
- To retry after a dropped stream, start a new request. There is no resume semantic.
SDK compatibility
Any OpenAI-compatible SDK works by changing base_url; your code stays the same. Verified: the official openai Python and Node SDKs, plus most desktop clients and terminal coding tools that allow a custom endpoint.
For per-client setup, see Use your existing SDK or client.