Cursor

The Cursor editor, via its OpenAI base-URL override (no config file). Reaches: the open-weight chat models (gpt-oss, GLM, DeepSeek, Kimi, Nemotron), Claude (sonnet, opus, haiku, …), and GPT-5.6 (gpt-5.6-sol / -terra / -luna) — the gateway's translation lanes bridge Cursor's wire to each model's native API. Works in Chat/Ask, Cmd-K, and Agent mode; Tab autocomplete always uses Cursor's own models.

Install: download the app at cursor.com.

A paid Cursor plan is required. Cursor's free (Hobby) plan only allows its "Auto" model, which ignores custom endpoints — selecting any named model shows "Named models unavailable — Free plans can only use Auto." Custom API keys work on Cursor Pro and above.

Cursor → Settings → Models

Turn on Override OpenAI Base URL and paste this — Cursor appends /chat/completions, so it must end in /v1:

https://app.inferencehub.tech/v1

Paste your key into the OpenAI API Key field:

sk-prov-live-YOUR_KEY

Add your models

Click Add model, enter an id exactly as your portal's Connect → Cursor tab lists it, and hit Verify. For example: gpt-oss-120b, zai-glm-5.2, sonnet, or gpt-5.6-terra. The flat-priced Auto routing ids work here too: add ih-auto or ih-auto-smart like any other custom model — unrelated to Cursor's built-in "Auto", which never uses custom endpoints. So do the Model synthesis ids (ih-synthesis, -budget, -max): each fresh prompt is answered by a panel of models and merged — expect a longer wait for the first token on those turns, while tool steps in Agent mode run at normal single-model speed.

Use ids exactly as listed. Cursor hijacks names it recognizes to its own providers: for Claude, use the short aliases (sonnet, opus) — never a claude-* name — and for open-weight models use the vendor-prefixed ids where listed (zai-glm-5.2, moonshotai-kimi-k2.6). A hijacked name fails with "Unauthorized User API key".

Pick the model and go

Select your custom model in Chat/Ask, Cmd-K, or Agent mode — requests route through InferenceHub and your spend appears in the portal under Usage & Billing.

Cost tip: Claude served through Cursor's OpenAI wire earns the prompt-cache read discount automatically — long agent sessions replay their context at cached-read rates, same as on the Anthropic wire. GPT-5.6 through Cursor also caches (automatic prefix caching).

Troubleshooting

If you see…	Fix
"Named models unavailable — Free plans can only use Auto"	Cursor's free plan blocks named-model selection, custom keys included. Upgrade to Cursor Pro (or use a different agent — Claude Code has no such gate).
`Unauthorized User API key` on a kimi / glm model	Cursor hijacked a bare model name to its own provider. Use the vendor-prefixed id (`moonshotai-kimi-k2.6`, `zai-glm-5.2`).
A `claude-*` model won't route to InferenceHub	Cursor hijacks `claude-*` names to its own Anthropic provider. Add the short alias exactly as listed (`sonnet`, `opus`, `haiku`).
Verify fails on every model	Check the base URL ends in `/v1` and the key is the full `sk-prov-live-…` value with no trailing space.