Choose your surface
One gateway, four common ways in. Pick by what your client speaks and whether you need one provider or many.
Quick pick
| If you are… | Use | Base URL | Model id shape |
|---|---|---|---|
| Writing an app/agent that may call OpenAI, Anthropic, Groq, Gemini, … from one client | Unified API | …/gateway/v1 | provider/model e.g. openai/gpt-4o-mini, anthropic/claude-sonnet-4-6 |
| Using the Anthropic SDK, Claude Code, or Claude Agent SDK | Anthropic passthrough | …/gateway/anthropic | Bare Anthropic id e.g. claude-sonnet-4-6 |
| Using Cursor Ask/Plan with Override OpenAI Base URL | Cursor adapter | …/gateway/cursor | Gateway ids e.g. openai/gpt-4o-mini |
| Pinning OpenAI only (OpenAI SDK, Agents SDK, Agno OpenAIChat, ADK→OpenAI) | OpenAI passthrough | …/gateway/openai/v1 | Bare OpenAI id e.g. gpt-4o-mini |
| Governing shell / MCP / prompts in the IDE (not model routing) | Relay | — | — → Relay Connect |
Gateway origin: https://production-api.exemplar.dev/gateway.
Default for most SDKs and frameworks: Unified API + Chat Completions (POST /gateway/v1/chat/completions) + eis_* key. Reach for passthrough or the Cursor adapter only when the client’s wire format requires it.
Side-by-side
Unified /gateway/v1 | Passthrough /gateway/{provider}… | Cursor /gateway/cursor | |
|---|---|---|---|
| Best for | Multi-provider apps, LangChain, Vercel AI, CrewAI, ADK with provider/model | Native Anthropic / OpenAI / Gemini clients that must keep vendor request shapes | Cursor IDE Override Base URL |
| Auth | eis_* | eis_* | eis_* |
| Model ids | openai/…, anthropic/…, … | Bare vendor ids (or LiteLLM’s own prefix rules) | Allowlisted gateway ids |
| Failover routes | Yes (Console Management → Routes) | No (prefix resolve only) | Uses unified Responses/Chat under the hood |
| Analytics / rate limits / cache | Yes | Rate limits yes; cache/failover differ by path | Yes (as gateway chat/responses) |
| Docs | Unified API | Passthrough | Cursor IDE |
Decision flow
Common mistakes
| Mistake | What to do instead |
|---|---|
Point Cursor at /gateway/v1 or /gateway/openai/v1 | Use /gateway/cursor — Cursor sends Responses-shaped bodies to /chat/completions |
Point Claude Code at /gateway/v1 | Use /gateway/anthropic with bare claude-… ids |
Use anthropic/claude-… on Anthropic passthrough | Drop the anthropic/ prefix on passthrough |
Use bare gpt-4o-mini on Unified when you need routing clarity | Prefer openai/gpt-4o-mini on /gateway/v1 |
| Expect Relay deny rules from the gateway base URL | Install Relay separately for tool governance |
Where to go next
Last updated on