A local AI that writes like you.

Privacy

Privacy model

• Ingestion, training, and generation are all local. No telemetry, no analytics, no account.
• The update check is the only routine network call. Sparkle requests one file — the release feed — and sends no system profile.
• Model weights and optional tools download on request. Weights come from Hugging Face on demand; the optional Messages tool is checksum-verified and fetched only when you click.
• The source is the proof. The repository is public so none of this has to be taken on faith.

Process

How it works

1. Ingest what you've already written — Mail, Messages, documents, chat exports.
2. Curate it into a personal training corpus, entirely on your Mac.
3. Fine-tune a local language model — QLoRA, via MLX — on your voice.
4. Compose — draft or rewrite text that sounds like you.

In use

A real Mac app, not a chat box.

Questions

Frequently asked

• Why not just use a SKILL.md or style-guide prompt instead? People write these to get an AI to sound more like them, sometimes specifically to slip past AI-detection tools. It can shift surface style — word choice, sentence rhythm — but the statistical signature of AI-generated text is still underneath it. Fine-tuning changes the model's actual weights on your writing, not just its instructions, which is a different and more durable change.
• Does it need an internet connection? No, other than checking for app updates (one file, no data sent) and any downloads you start yourself — model weights, the optional Messages tool. Ingestion, training, and generation all run offline, on your Mac.
• How much of my writing does it need? There's no hard minimum, but more of it — and more variety, across different audiences and moods — produces a better fine-tune. The app shows your corpus size and lets you retrain as it grows.
• Can it just regurgitate my messages verbatim? Training includes a check for verbatim overlap between generated output and your source text, so memorization gets flagged rather than passed off as a draft.
• Does it end up training on its own output? Not by accident — a self-ingestion guard keeps generated drafts out of the training corpus by default, so it doesn't quietly drift toward its own phrasing. But if you edit a draft and keep it, that edited version really is your writing at that point, and you can save it back in to improve future fine-tunes.
• Does it work on Windows, Linux, or Intel Macs? No — Apple Silicon and macOS only. That's also what makes iMessage such a good source here: it's already sitting on your Mac in bulk, in your actual voice, with no export or account access needed.
• Is this open source? The source is public and auditable under a source-available license (PolyForm Noncommercial) — free for personal use; commercial use requires a separate license.
• Does it cost anything? Writekin is free for personal, noncommercial use under the PolyForm Noncommercial 1.0.0 license. Using it in a product, service, or other commercial context requires a separate commercial license — email scottjg@gmail.com to arrange one.

Requirements

Requirements & install

• Apple Silicon — 16 GB unified memory minimum, 32 GB+ recommended for training larger models.
• macOS 14 or newer.
• Full Disk Access — optional, user-granted, only if you want Mail and Messages ingested.

Messages support uses the open-source imessage-exporter (GPL-3.0), downloaded on demand — never bundled with the app.