Grammar checkers see your most sensitive text: unreleased docs, client contracts, internal incident reports. Before you paste that into a browser extension, ask the question most people skip: where does this text actually go?
The Three Privacy Models
| Model | How it works | Who sees your text |
|---|---|---|
| Cloud SaaS | Text goes to vendor servers | The vendor (and possibly subprocessors) |
| BYOK | Text goes through an LLM API with your key | The model provider you already chose |
| Self-hosted / local | Everything runs on your machine | Nobody |
What to Check in a Vendor's Policy
- Retention - is your text stored after the check? Used for "service improvement" (training)?
- Training opt-out - opt-out defaults mean your drafts train their models unless you find a settings page
- Subprocessors - a "we don't store your data" promise means little if an inference provider does
- Compliance - NDA'd client data or regulated text may disqualify cloud SaaS outright
Where BYOK Fits
BYOK is the middle path that suits developers: the tool UI never holds your text in a vendor database - requests go straight to the model provider under your account. Lint works this way: connect any OpenAI-compatible key (including DeepSeek) and usage is unlimited and free. Your text goes to the provider you picked, and Lint keeps nothing.
When You Truly Need Self-Hosted
Regulated industries and air-gapped environments justify running something like LanguageTool locally. You trade style accuracy for total control. For most developers, a BYOK setup delivers 95% of the privacy benefit with none of the ops burden.
Match the tool to the sensitivity of the text: throwaway drafts anywhere; client work behind BYOK; regulated text selfhosted.
👉 Try Lint's BYOK mode - your key, your provider, unlimited free checks.
Top comments (0)