fine-tune in the browser
Pick a model, feed it messages, and train with a live loss chart, sane defaults, and one-click load-and-chat when it's done.
the cover artist
Reprise
A fully offline pipeline that fine-tunes a local MLX model on your own iMessage history â with privacy redaction and a human review step â until it texts like you. Everything runs on the Mac; nothing is uploaded anywhere, ever.
Python
MLX fine-tuning
fully offline
iMessage
What if your computer could sing your songs? Reprise learns your texting voice â the phrasing, the languages you mix, the way you actually reply â by fine-tuning a local MLX model on your own Messages history. It's a cover artist trained on the original recordings, and the recordings never leave the studio: the whole pipeline runs offline on one Mac.
Privacy isn't a feature here, it's the architecture. The iMessage pipeline redacts sensitive content and puts a human review step between your history and any dataset. Even the AI coding assistants that work on the repo are barred, by hard rule, from ever reading message content â they only see counts and aggregates.
The workflow is four pieces: a Studio web app where you pick a base model, build a dataset, watch the loss curve, and chat with the result; a dataset pipeline; a one-click Extractor app; and a Relay app that lets the model answer real iMessages â with a floating STOP button, because of course there's a floating STOP button.
highlights
iMessage pipeline
Turns your Messages history into a training dataset with redaction, language filtering, and a human review step.
Extractor.app
A double-clickable Finder app that refreshes your Messages data in one step â no terminal required.
Relay.app
Lets Reprise answer real iMessages for you, with a status panel and a floating STOP button always within reach.
offline by design
Vendored wheels mean the whole environment reinstalls with zero internet. Nothing is uploaded anywhere, ever.
probe before you train
A one-minute real training test tells you whether a base model will fine-tune before you commit hours to it.
under the hood
- mlx-lm fine-tuning â LoRA/DoRA training on any compatible local MLX model, with live train and validation loss.
- Python 3.12 + FastAPI â the Studio is a local web app on 127.0.0.1; models are discovered from local caches.
- redaction first â sensitive content is stripped and a review report generated before any dataset is finalized.
- multilingual datasets â language filters bake a matching respond-only-in instruction straight into training.
- evals included â vibe-check evals and memorization probes keep an overfit model from ever reaching the relay.