Provider integration

Use any OpenAI-compatible endpoint in your coding agents

If it serves OpenAI Chat Completions or Responses, AutoJev can route to it. Add the base URL and key once, and Codex, Claude Code, Gemini CLI, OpenCode and the rest can all use it.

OpenAI Compatible preset with a custom base URL
Chat Completions or Responses, set per model
Hosted APIs and local servers alike
Built-in provider test shows the exact endpoint

Download AutoJev Star on GitHub

Providers

+ Providers

OpenRouter

https://openrouter.ai/api

DeepSeek

https://api.deepseek.com

OpenAI

https://api.openai.com/v1

Ollama

http://127.0.0.1:11434

One preset for the long tail of providers

Most model providers and local inference servers now expose an OpenAI-style API. AutoJev does not need a dedicated preset for each of them: the OpenAI Compatible preset takes any base URL and API key, and you choose the upstream API type per model.

Once added, those models behave like any other in AutoJev. They can be connected to agents directly as `<provider>/<model_id>`, grouped into routes, load-balanced, speed-tested and logged. Agents that speak a different protocol — Codex with Responses, Claude Code with Anthropic Messages — are converted automatically.

Add an OpenAI-compatible provider

1
Choose OpenAI Compatible
Providers → Add provider → OpenAI Compatible (under Custom). Give it a short provider name such as `kimi` or `vllm`; it becomes the prefix of the model IDs agents see.
text
Provider type: OpenAI Compatible Provider name: kimi Base URL: https://api.moonshot.ai/v1 API type: OpenAI Chat Completions API key: sk-…
2
Get the base URL right
AutoJev appends /v1/chat/completions (or /v1/responses) to the base URL. If the base URL already ends in /v1, it is not duplicated. An endpoint whose OpenAI path does not end in /v1 will not resolve correctly — use the provider test to check.
3
Test
Enter a model ID and click Test. The dialog shows the exact test endpoint, for example https://api.moonshot.ai/v1/chat/completions, and the upstream error if anything is wrong.
4
Add models and connect
Add each model with its ID, API type, prices, tier, capability flags and context window. Then select it on the Agents page, or put it in a route.

Endpoints people commonly add

Any OpenAI-compatible endpoint works, such as the ones below. Base URLs are taken from each provider’s documentation; regions and keys are separate, so use the URL that matches where your key was issued.

Provider or server	Base URL
Kimi / Moonshot (global)	https://api.moonshot.ai/v1
Kimi / Moonshot (China)	https://api.moonshot.cn/v1
Qwen / DashScope (China)	https://dashscope.aliyuncs.com/compatible-mode/v1
Qwen / DashScope (international)	https://dashscope-intl.aliyuncs.com/compatible-mode/v1
SiliconFlow (China / global)	https://api.siliconflow.cn/v1 · https://api.siliconflow.com/v1
MiniMax	https://api.minimax.io/v1
vLLM (default local server)	http://localhost:8000/v1
LM Studio (default local server)	http://localhost:1234/v1

Providers change URLs and regions over time; check their docs if a test fails. GLM / Zhipu’s OpenAI-style path does not end in /v1, so add GLM through its Anthropic-compatible endpoint instead.

Use it from any client

Besides one-click agent connections, the Connection parameters dialog gives you env, JSON and cURL snippets for any BYOK client. The key is a local placeholder; an optional client name becomes its suffix so requests are labeled in Logs.

Connection parameters — OpenAI protocol, env format

OPENAI_BASE_URL=http://127.0.0.1:9527/v1
OPENAI_API_KEY=autojev-local-my-script
# then request model "kimi/<model_id>" or a route such as "autojev/fast"

What AutoJev adds to a compatible endpoint

Agent compatibility

Codex and Claude Code can use Chat-only providers through Responses and Messages conversion.

Resilience

Put two compatible providers in one route. Failover retries the next candidate on connection errors, 429 or 5xx before output starts.

Comparable numbers

Speed tests, logs and usage use the same metrics for every provider, so you can compare a self-hosted vLLM with a hosted API.

Compatible is not identical

Providers differ in which OpenAI features they implement. Set capability flags honestly; conversion cannot add tools, vision or reasoning that the upstream model lacks, and provider-specific fields are only preserved on native (unconverted) requests.

Frequently asked questions

Which providers can I add with the OpenAI Compatible preset?

Any endpoint that implements OpenAI Chat Completions or Responses, such as Kimi/Moonshot, Qwen on DashScope, SiliconFlow, MiniMax, or local servers like vLLM and LM Studio.

Should my base URL include /v1?

Either works when the provider’s path is /v1/…: AutoJev appends /v1/chat/completions, and skips the extra /v1 if the base URL already ends in it. The provider test shows the final URL.

Can I use GLM / Zhipu models?

GLM’s OpenAI-style endpoint uses a path that does not end in /v1, which AutoJev cannot address directly. Add GLM through its Anthropic-compatible endpoint with the Anthropic Compatible preset instead.

Do local servers like vLLM or LM Studio need an API key?

That depends on how you started the server. If it does not check keys, any placeholder value works; if it does, enter the key you configured.

Can Codex use a provider that only supports Chat Completions?

Yes. Codex sends Responses to AutoJev, which converts to Chat Completions and back, including streaming and function tools.