Skip to content

AI

AI is quietly flattening everyone's voice.

Code Video Cerebral Valley

Core trick

Distills model labels into three local classifiers that filter a personal writing corpus, then serves the fine-tuned Ollama model as an MCP server.

Tags

productivitydev tools

Description

AI is quietly flattening everyone's voice. The more I used Claude and ChatGPT to write faster, the less I sounded like myself. Kiln is a Mac app that fine-tunes a local LLM on your own writing in 20 minutes. You drop a folder, or connect Notion and Apple Notes through MCP. An Opus 4.7 orchestrator spawns sub-agents in parallel to clean your corpus. Three classifiers we distilled from 5,000 Opus labels run locally to judge quality, preference, and style. Once trained, your voice runs on Ollama. And the cool part: Kiln also exposes it as an MCP server. So you can plug it into Claude.app and Claude will write as you.
timothim/kiln★ 01 forksSwiftMITCreated Apr 22, 2026 · last push Apr 27, 2026

README excerpt

<div align="center">

```
    ██╗  ██╗██╗██╗     ███╗   ██╗
    ██║ ██╔╝██║██║     ████╗  ██║
    █████╔╝ ██║██║     ██╔██╗ ██║
    ██╔═██╗ ██║██║     ██║╚██╗██║
    ██║  ██╗██║███████╗██║ ╚████║
    ╚═╝  ╚═╝╚═╝╚══════╝╚═╝  ╚═══╝
```

**Train AI to write in your voice. Local. Private. Yours.**

[![Built with Opus 4.7](https://img.shields.io/badge/Built_with-Opus_4.7-D97706?style=flat)](https://www.anthropic.com)
[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
[![Swift 5.9+](https://img.shields.io/badge/Swift-5.9+-orange.svg)](https://swift.org)
[![Python 3.11+](https://img.shields.io/badge/Python-3.11+-3776AB.svg)](https://www.python.org)
[![macOS 14+](https://img.shields.io/badge/macOS-14+-000000.svg)](https://www.apple.com/macos)

</div>

Kiln is a native macOS app that fine-tunes a small local LLM on your own writing in about twenty minutes on a MacBook. Drop a folder of your emails, notes, messages, and Markdown — Kiln dedupes the corpus, scores it with three Opus-distilled classifiers running on-device, runs LoRA fine-tuning via MLX, and exposes the trained voice through Ollama and as an MCP server. Your data never leaves the laptop. Anthropic's Opus 4.7 was used at development time to teach three small classifiers Kiln ships locally, and is available as opt-in cloud advisor features (Voice Coach, Training Advisor, Deep Curation).

> **Submission for** the Built with Opus 4.7 hackathon (April 21–26, 2026). [**▶ Watch the 3-minute demo on You

Team