Writekin: Fine-tune local LLM on your Mac with your own writing

Writekin is a macOS application that fine-tunes a local language model on a user's own writing to generate and rewrite text in the user's voice. The app ingests writing from Mail, Messages, and documents, curates the text into a training corpus, then runs QLoRA (a parameter-efficient fine-tuning method) on Apple Silicon hardware via MLX, and finally uses the trained model to draft and rewrite in the user's style. All processing—ingestion, training, and generation—remains on the user's Mac; nothing is sent to external servers.

The privacy model is strict by design. The app sends no telemetry or analytics and requires no user account. The only network traffic is an auto-update check (via Sparkle) that requests a release feed and sends no system profile data. Model weights download from Hugging Face on demand when the user clicks to fetch them. If the user enables Messages support, Writekin downloads an open-source tool called imessage-exporter from GitHub (with checksum verification) only when explicitly initiated. The developer publishes the full source code so users can audit the codebase and verify these claims without relying on trust alone.

System requirements are modest by modern standards but not trivial: Apple Silicon Macs (M-series chips) with 16 GB of unified memory minimum, and 32 GB or more is recommended for training larger models. macOS 14 or newer is required. To ingest Mail and Messages, the app needs Full Disk Access, which the user must grant explicitly. Installation is straightforward—download the latest DMG file from the Releases page, drag it to Applications, and open it; the app guides the user through source detection, Full Disk Access, and model download on first launch. Users can also build from source using xcodegen and Xcode.

Licensing is tiered: Writekin is source-available and free for personal noncommercial use under the PolyForm Noncommercial 1.0.0 license. Commercial use requires a separate license, with contact information in NOTICE.md. Third-party component licenses are documented in THIRD_PARTY_LICENSES.md. The codebase includes a test suite (xcodebuild test -scheme Writekin) that must pass before changes are accepted, and newcomers are directed to ARCHITECTURE.md for an overview of the pipeline, folder structure, and design invariants.

Writekin addresses a specific gap in the AI writing-assistance market: users who want personalization without surrendering their writing data to cloud providers. By running the entire pipeline—ingestion, fine-tuning, and inference—on-device using Apple Silicon, the app avoids the privacy tradeoff many cloud-based writing tools impose. The architecture uses QLoRA (Quantized Low-Rank Adaptation), a parameter-efficient fine-tuning technique that reduces memory and compute requirements, paired with MLX, an Apple-optimized machine learning framework. This technical choice makes the task feasible on consumer hardware (16–32 GB RAM) rather than requiring expensive GPUs or cloud subscriptions.

The transparency model—publishing source code and documenting the architecture—is a deliberate choice to let users verify privacy claims rather than rely on the developer's assertions. Model weights come from Hugging Face, and optional integrations (like Messages export) use open-source tools (imessage-exporter) with checksum verification, lowering the risk of supply-chain compromise. The licensing split (free personal, paid commercial) targets hobbyist and professional writers separately, positioning Writekin as a tool for users who value control over convenience.

No comments yet. Be the first to share your thoughts!