Skip to main content
ByteSpike exposes a small, well-defined set of endpoints behind one API key. Each one matches a different client persona — an existing SDK, an agent framework, a multimodal pipeline. The gateway translates between protocols so you don’t write per-vendor glue. Principle: pick the protocol your client already speaks; the gateway handles the rest.

​
The endpoint surface

That’s it. No /v1/embeddings, no /v1/audio/*, no /v1/rerank, no /v1/assistants — ByteSpike does not currently expose those surfaces.

​
1. Anthropic Messages — /v1/messages

The canonical agent protocol on ByteSpike. Preserves tool_use, cache_control, thinking, and web_search blocks end-to-end. Use when:
  • You’re building an agent that needs explicit tool-call / tool-result blocks
  • You want extended thinking on Opus / Sonnet 4.x
  • Your client is the official anthropic SDK (Python / Node) or a derivative
Works with Claude models natively, and with DeepSeek, Gemini, and GPT models via transparent translation — pass any catalog model in the model field. See /v1/messages reference.

​
2. OpenAI Responses — /v1/responses

The newer OpenAI agent-style protocol (Codex CLI + Agents SDK). Use when:
  • Your client is OpenAI Codex CLI
  • You want structured-outputs + reasoning combined
  • You’re on the OpenAI Agents SDK
Works with GPT models natively, and with Claude and Gemini via transparent translation. See /v1/responses reference.

​
3. OpenAI Chat Completions — /v1/chat/completions

The most ubiquitous chat shape in the ecosystem. Almost every SDK speaks it. Use when:
  • You have existing openai SDK code and want the lowest-friction route
  • You’re not writing an agent — chat in, chat out
  • You’re calling a non-OpenAI model but want to keep the OpenAI SDK
Works with GPT, Gemini, DeepSeek, Kimi (Moonshot), GLM (Zhipu), MiniMax, and Doubao (ByteDance) models — pass any catalog model in the model field. See /v1/chat/completions reference.

​
4. Gemini Native — /v1beta/models/{model}:generateContent

Google’s native protocol verbatim. Use when:
  • You’re using Google’s official CLI or SDK
  • You want grounding (googleSearch tool) in its native form
  • You need Gemini’s exact response shape for downstream tooling
Streaming via :streamGenerateContent?alt=sse. This protocol serves the Gemini family only. See /v1beta/... reference.

​
5. Image generation (sync) — /v1/images/generations + /v1/images/edits

OpenAI-shape body across providers; the model field routes. Use for:
  • Single-prompt image generation (GPT-Image-2 shape)
  • Image edits with a mask or instruction
Serves the GPT Image 2 (OpenAI) and Nano Banana (Google) image models — all through the same OpenAI-shape body.

​
6. Async multimodal tasks — /v1/tasks/{submit, query, cancel, stream}

Video and long-running image generation use the async task model: submit returns a task_id immediately, poll /tasks/query or stream state changes over SSE on /tasks/stream/{task_id}. Use for:
  • Video generation (Veo 3.1) — takes 10-60s+
  • Batched image generation where you want fire-and-forget
  • High-concurrency producers that shouldn’t hold HTTP connections open
See /tasks/submit reference.

​
Which protocol for which model family

Which protocol you can call each model family through on ByteSpike. ✅ = supported; — = not available on that protocol. For the live, account-specific list, hit GET /api/v1/me/available-models or click Test next to any model in the console.

​
Picking by client

​
Scenarios

​
One key, multi-endpoint

A ByteSpike API key today binds to one routing group (group_id). The group determines which model families the key can reach. Within a group, every endpoint shape those models support is available — e.g. a key in claude-default can call /v1/messages, /v1/chat/completions (translated), and /v1/responses (translated) against Claude models. To reach model families in different groups, create multiple keys (one per group) and select per request. See Authentication.

​
Next