Supported BYOK providers and models
The AI providers Skrebe supports, the models you can pick per provider, and how your key is stored.
Skrebe runs AI features — translation and proofread — through one of three providers. You choose the provider, you bring the API key. This page is the lookup table.
At a glance
| Provider | Models | Where to get a key |
|---|---|---|
| Google Gemini | Gemini 3.1 Flash Lite · Gemini 3.5 Flash | aistudio.google.com |
| Anthropic Claude | Claude Opus 4.8 · Claude Sonnet 4.6 · Claude Haiku 4.5 | console.anthropic.com |
| OpenAI GPT | GPT-5.5 · GPT-4.1 · GPT-4.1 Mini | platform.openai.com |
Skrebe defaults to Google Gemini 3.1 Flash Lite because it's roughly 4× faster and 6× cheaper than Gemini 3.5 Flash for short paragraph-sized rewrites, which is what Skrebe asks the model to do most of the time — and it holds up just as well on quality.
Picking a model
A few rules of thumb — Gemini keeps it simple with two tiers instead of three:
- Opus / Sonnet / GPT-5.5 for long documents, tricky idioms, or when you've reviewed the output and want a second pass with more nuance.
- Flash / Haiku / GPT-4.1 for routine drafting and quick paragraph polishing. Solid all-rounders for Kriolu (ALUPEC) drafting.
- Flash Lite / GPT-4.1 Mini — the cheapest, fastest tier. Gemini's Flash Lite holds up well here — it's Skrebe's own default. On GPT-4.1 Mini, give nuanced passages a quick read-through before you ship.
Each provider ships with its own default when you first pick it: Gemini defaults to Flash Lite, Claude defaults to Sonnet 4.6, GPT defaults to GPT-4.1 — so "the default" isn't the same tier across providers. You can switch model at any time from /settings without re-entering your key.
Storage details
Your key is read on-page and held in memory while Skrebe is open. Reloads pull it back from local storage. Cleared site data → key gone. Skrebe never logs the key, never sends it to a Skrebe-owned server, and never sends it cross-provider.
Switching providers
From the /settings page, change the provider dropdown, paste the new key, hit save. Press Esc to dismiss the modal. The next translation or proofread request uses the new provider.
What Skrebe does NOT support yet
- Local models (Ollama, llama.cpp, etc.) — tracked for a future spec.
- Custom endpoint URLs (proxies, Azure OpenAI, Vertex AI gateways).
- Per-feature provider overrides (e.g. "use Claude for translation, Gemini for proofread"). One provider, one model, app-wide.
For free-tier limits and BYOK quota behavior, see Free-tier rate limits + BYOK quotas.