Please Added support for LLM Endpoint

[Doutianbao]

Thank you for providing such a great open source project. Can you consider adding support for otheropensource llm proxy projects? such as sub2api or newapi?
Sub2api :https://github.com/Wei-Shaw/sub2api

[TKassis]

Thanks for the kind words — and good news: this already works today, no new code needed. It just isn't obvious, because the env var you want is the OpenRouter one rather than the one named "OpenAI-compatible".

Pointing Kady at sub2api / new-api

In the repo-root .env:

OPENROUTER_BASE_URL=http://your-proxy:3000/v1
OPENROUTER_API_KEY=<your proxy token>

That path sends your key as a Bearer token against a plain /v1/chat/completions endpoint, which is exactly what these proxies expose. Usage still counts against the project spend cap.

Two caveats worth knowing:

• The model picker lists ids from the bundled OpenRouter catalogue (web/src/data/models.json). If your proxy names models differently, they won't show up in the dropdown — you can still use them with DEFAULT_MODEL_PROVIDER=openrouter and DEFAULT_MODEL_ID=<your id>.
• Cost is looked up from that catalogue by model id, so an id we don't recognise gets ledgered at $0. The run works fine; the spend number is just wrong for it.

To be straight with you: this route was built and tested against gateways that mirror OpenRouter's vendor/model ids and Bearer auth (Requesty is the case the code comment names). sub2api and new-api are OpenAI-shaped in the same way, so it should work, but neither is something we've verified end to end. If you try it, please report back — happy to fix whatever doesn't line up.

Why not OPENAI_COMPATIBLE_BASE_URL

That's the variable you'd expect to reach for, and it won't work here. It was built for local servers (LM Studio, vLLM) that need no credential, so:

• no API key is ever sent — the provider is registered with a placeholder credential and model discovery fetches /v1/models with no Authorization header, so an authenticated proxy just 401s and the picker section stays hidden;
• the provider is hardcoded as local/$0 in the billing code, so paid spend through it would be untracked and uncapped.

Also worth checking first

If your reason for wanting sub2api is turning a Claude or ChatGPT subscription into an API, Kady does that natively already — Settings → Model providers has OAuth for Anthropic (Claude Pro/Max), OpenAI Codex (ChatGPT Plus/Pro), GitHub Copilot and xAI, no proxy in between.

Leaving this open as a feature request for proper first-class support of authenticated OpenAI-compatible gateways, which would need: an API key for the openai-compatible provider, a billing mode for it other than local/$0, and ideally more than one configurable endpoint.

[Doutianbao]

Thank you for your reply. I followed your recommendation and used the alternative URL, key (which I set as "gpt-5.6-terra"), and model. However, after launching, no matter which "gpt-5.6-terra" option I select in the dialog box, it doesn't work properly.

[Doutianbao]

Another question is, when integrating with an OpenAI-compatible provider in the future, do I also need to take the "thinking level" parameter into consideration?