Promptshot now supports three model backends. They are not equivalent, and the marketing-friendly answer — “use whichever you prefer” — is not useful when the differences are this concrete. Here is the actual trade.
Hosted (the default)
You do nothing. Requests go to our endpoint, which routes to a small fast model.
- Latency: 280–420 ms end to end for a two-sentence rewrite
- Setup: none
- Cost: included in the licence
- Privacy: the selected text transits our infrastructure and is not retained after the response
This is the right default for the same reason a fast car is the right default when the journey is short: for a forty-word Slack message, the difference between a small model and a frontier model is barely visible, and the difference in latency is very visible.
Your own API key
Enter a key under Settings → Model → Bring your own key. Requests go directly from your Mac to the provider; nothing passes through us.
Provider: Anthropic
Model: claude-sonnet-5
Key: sk-ant-... (stored in the macOS Keychain)- Latency: 500–900 ms, depending on provider and model
- Setup: two minutes
- Cost: billed by the provider, typically well under a dollar a month at ordinary usage
- Privacy: governed by your own account terms, which for most business tiers means no training on your data
Use this if you need a frontier model for long-form editing or translation, or if your organisation requires a direct contractual relationship with the model provider. The latency cost is real but tolerable for edits you make a few times a day rather than forty.
Local
Point Promptshot at any OpenAI-compatible endpoint on localhost — Ollama, LM Studio, llama.cpp’s server.
Endpoint: http://127.0.0.1:11434/v1
Model: qwen3:8b- Latency: 600 ms to 4 s, entirely dependent on your hardware and model size
- Setup: fifteen minutes plus a model download
- Cost: electricity
- Privacy: absolute; the text never leaves the machine
We support this because some people genuinely cannot send text off-device, and for them the latency is not a trade-off, it is the only option. We would be doing you a disservice to pretend it is otherwise fast. On an M4 Pro with an 8B model, a short rewrite takes about 900 ms — noticeably slower than hosted, still usable. On an M1 Air with anything larger than 8B, it is slow enough that you will stop using the hotkey.
If you are choosing local for privacy reasons, prefer a smaller model and a tighter prompt over a larger model. Article and preposition correction does not need 30B parameters, and the latency difference decides whether the feature gets used at all.
Mixing them
Model choice is per profile, so the three backends compose:
| Profile | Backend | Why |
|---|---|---|
| Polish | Hosted | Forty times a day, latency-critical |
| Translate | API key | Frontier model handles source grammar better |
| Confidential | Local | Contract text that cannot leave the device |
That last row is the reason we built all three rather than picking one. Most people need speed for most of their writing and privacy for a little of it, and forcing a single global choice means compromising on both.
Available now in 1.4.