SDK Integration Overview

The gateway is a protocol-oriented entry point: whatever SDK you use on the client, as long as it speaks one of the gateway's supported protocols (OpenAI / Anthropic / Gemini / DashScope / Realtime / MCP / passthrough), pointing its base_url at the gateway and swapping its API key for a gateway-issued access key lets you call any model configured in the gateway — including cross-protocol calls to other vendors' models. This group of pages is organized by "which SDK you already have", each with a copy-paste minimal example.

The three universal steps

Every SDK follows the same wiring rule:

1. Point base_url at the gateway (default http://localhost:7890). Note that each SDK differs on whether base_url includes /v1 — see the per-page examples.
2. Set the API key to a gateway-issued access key, not an upstream sk-.... Access keys are issued on the Console's "Access keys" page.
3. Set model to the model name configured in the gateway. The gateway does not restrict calls by model name — whatever name you configure in the console is what the client uses.

Model names that appear in examples (gpt-5.4-mini, claude-haiku-4-5, gemini-3.5-flash, qwen3.8-max, …) are just "real names under some gateway config"; use whatever your console defines.

Auth headers

Data-plane endpoints try the following auth methods in order, first match wins (see Endpoints · Auth):

Header	Corresponding SDK	Notes
Authorization: Bearer <access key>	OpenAI, DashScope, generic	Bearer prefix is case-insensitive
x-api-key: <access key>	Anthropic SDK	Authorization also accepted
x-goog-api-key: <access key>	Gemini SDK	Authorization also accepted
(none of the above)	—	Anonymous fallback, only if an anonymous=true key is configured

The three headers are just different mounting slots for the same "gateway access key" — issue one in the console and put it in whichever header your SDK expects.

SDK selection matrix

"I'm already using X — how do I connect it to the gateway?" Find your row:

SDK / framework	Language	Protocol / endpoint	Text	Image	Video	Realtime	Embed/rerank	Page
OpenAI SDK	Py / Node	openai family (/v1/chat/completions, /v1/responses, …)	✅	✅	—	✅¹	✅	Vendor SDKs
Anthropic SDK	Py / Node	anthropic (/v1/messages)	✅	input understanding	—	—	—	Vendor SDKs
Google GenAI SDK	Py / JS	google (/v1beta/models/*)	✅	✅	—	—	—	Vendor SDKs
DashScope SDK / HTTP	Py	dashscope (/v1/services/*)	✅	✅	✅ async	✅¹	✅	DashScope native & async
OpenAI-compatible third parties (DeepSeek/GLM/Kimi/Grok…)	Py / Node	openai compatible	✅	vendor-dependent	—	—	vendor-dependent	Vendor SDKs
Vercel AI SDK	TS	openai / anthropic (provider layer)	✅	provider-dependent	—	—	provider-dependent	Unified SDKs & gateways
LiteLLM	Py	openai compatible	✅	model-dependent	—	—	✅	Unified SDKs & gateways
LangChain / LangGraph	Py / TS	underlying chat model (OpenAI/Anthropic…)	✅	model-dependent	—	—	✅	Orchestration frameworks
LlamaIndex	Py / TS	OpenAILike / chat model	✅	model-dependent	—	—	✅	Orchestration frameworks
CrewAI	Py	underlying LiteLLM	✅	model-dependent	—	—	—	Orchestration frameworks
AutoGen / AG2	Py	openai compatible	✅	model-dependent	—	—	—	Orchestration frameworks
OpenAI Agents SDK	Py / TS	openai_response (default)	✅	model-dependent	—	—	—	Agent SDKs
Claude Agent SDK	Py / TS	anthropic	✅	input understanding	—	—	—	Agent SDKs
Pydantic AI	Py	openai compatible	✅	model-dependent	—	—	—	Agent SDKs
Realtime (WebSocket)	any WS client	openai_realtime (/v1/realtime)	—	—	—	✅	—	Realtime & audio

¹ Realtime connects through the gateway's OpenAI Realtime client protocol (/v1/realtime) over a raw WebSocket — no vendor SDK is required; model is the realtime model's config name in the gateway (OpenAI or DashScope).

Cross-SDK pitfalls

No matter which SDK, after going through the gateway watch for these (each page repeats them; here they are in one place):

• The model name must match the gateway config. The model you pass is "the name configured in the gateway" — not the vendor's official name, not the upstream's real model ID. Wrong name → 404 model_not_found.
• Whether base_url includes /v1 varies by SDK. OpenAI SDK takes /v1; Anthropic SDK takes the root (it appends /v1/messages); Google GenAI takes the root (via http_options.base_url). Getting it wrong adds or drops a path segment.
• Fields get translated and constrained by the target protocol across protocols. For example, if the gateway's gpt-5.6-sol upstream is a Responses protocol, a max_tokens passed by the client is translated to max_output_tokens, and the upstream may impose a lower bound (e.g. ≥16) — too small → 400.
• Thinking blocks / signatures are wrapped by the gateway. Claude and Gemini thinking content (thinking / thoughtSignature) is passed through or stripped by the gateway; to get the text, take the type == "text" element of the content array — don't assume content[0] is text.
• Third-party SDK model-name allowlists. Some frameworks (LlamaIndex, AutoGen, …) validate the model name against families they recognize and reject unknown gateway names — use their "any model" entry point or declare the model info explicitly (see each page).

Next

• Want the smallest possible call: start with Vendor SDKs.
• Frontend / Next.js: see Unified SDKs & gateways.
• Already on LangChain / LlamaIndex / CrewAI: see Orchestration frameworks.
• Voice / video and other niche capabilities: see Realtime & audio and DashScope native & async tasks.
• Full provider and model list (without SDK examples): see Supported providers & models.