Protocol compatibility
Keep your SDK’s protocol format and check support for each model.
Protocol compatibility lets you keep your client’s request format. It does not mean every model supports every protocol or every official feature. Check the model’s supported_apis and endpoint support notes.
One address, https://api.deepraxis.com, with one entry per protocol family:
| Entry | Protocols |
|---|---|
/openai | OpenAI protocols |
/anthropic | Anthropic protocols |
/google | Google Gemini API |
OpenAI Chat Completions
POST /openai/v1/chat/completions. A common chat protocol; streaming (SSE), tool calls, and multimodal input depend on the model. See Chat API for field support and session affinity for session identifiers.
OpenAI Responses
POST /openai/v1/responses plus retrieve / delete / cancel / input_items. OpenAI's stateful next-generation protocol; agents like Codex use it. See Responses API for support. Response compaction and input-token counting endpoints are not implemented and return 501.
OpenAI Files
/openai/v1/files: upload user_data files for later requests; list, retrieve metadata, download, delete.
OpenAI Images
POST /openai/v1/images/generations: use a model that declares image-generation support in the catalog.
Anthropic Messages
POST /anthropic/v1/messages. Claude Code and the Anthropic SDK use it. See Messages API for support.
Anthropic Files
/anthropic/v1/files: upload, list, retrieve metadata, delete; uploads are not downloadable (same as upstream).
Google Gemini API (generateContent)
POST /google/v1beta/models/{model}:generateContent. Currently image generation and editing with the native Gemini protocol.
Model catalog
Use GET /openai/v1/models?api=responses to find Responses models; api=chat,embeddings matches either API. The native catalog GET /model/v1/models and its model detail endpoint expose each model’s supported_apis; filter the native list with ?supported_api=openai.responses.
GET /openai/v1/models: the models this key can call, OpenAI-shaped; clients such as Kilo Code discover models through it.
Anthropic clients use GET /anthropic/v1/models; Google clients use GET /google/v1beta/models. Each returns the corresponding protocol’s model list.
Standardized paths
Every entry also has a standardized long path /v1/router/<vendor>/... with identical semantics — see the API reference.
→ Go deeper: OpenAI-compatible API · Anthropic-compatible API