On-device PII Detection & Redaction, Powered by MLX & ONNX Runtime.

Author: EvilChen

README Version: [English | 简体中文]

Project Introduction

Karma is a privacy-first desktop application for detecting and redacting Personally Identifiable Information (PII), based on OpenAI Privacy Filter model. All AI inference runs entirely on-device — your data never leaves your machine.

• macOS (Apple Silicon): Powered by MLX for native high-performance inference
• Windows / Linux: Powered by ONNX Runtime with automatic GPU acceleration (CUDA / DirectML / CPU fallback)

Karma identifies 8 types of PII in text and replaces them with labeled placeholders:

PII Type	Placeholder
Person Name	[PERSON]
Email Address	[EMAIL]
Phone Number	[PHONE]
Physical Address	[ADDRESS]
Date	[DATE]
URL	[URL]
Account Number	[ACCOUNT_NUMBER]
Secret / Password	[SECRET]

Features

• Real-time PII Detection: Token-level classification with platform-optimized inference
• Cross-platform: macOS (MLX) / Windows (ONNX + CUDA/DirectML) / Linux (ONNX + CUDA/CPU)
• Text Redaction: One-click redaction with auto-copy to clipboard
• HTTP API Server: Built-in REST API server (/health, /redact) for integration with other tools
• Redaction History: SQLite-backed history with infinite scroll browsing
• Global Shortcut: Quick access via configurable keyboard shortcut (default: Cmd+Shift+K)
• Bilingual UI: English and Simplified Chinese
• Zero Cloud Dependency: All processing happens locally for complete privacy

Tech Stack

Component	Technology
Desktop Framework	Tauri 2 (Rust)
Frontend	React + TypeScript + Tailwind CSS
State Management	Zustand
AI Inference (macOS)	MLX (via mlx-rs, in-process)
AI Inference (Windows/Linux)	ONNX Runtime (via ort, CUDA / DirectML / CPU)
HTTP Server	Axum
Database	SQLite (rusqlite)
Build Tool	Vite

Installation

Download the latest release from the Releases page:

Platform	File	Requirements
macOS	.dmg	macOS 15.0+ with Apple Silicon (M1/M2/M3/M4)
Windows	.exe	Windows 10+ (x64)
Linux	.deb	Ubuntu 22.04+ (x64)

Usage

Model Download

Platform	Model	Link
macOS (Apple Silicon)	MLX format	mlx-community/openai-privacy-filter-bf16
Windows / Linux	ONNX format	yasserrmd/privacy-filter-ONNX

Quick Start

1. Open Karma and navigate to the Model page
2. Select your local model directory (MLX or ONNX depending on your platform)
3. Wait for the model to load (status indicator turns green)
4. Switch to the Redactor page, paste your text, and click Redact

Global Shortcut

Press Cmd+Shift+K (configurable in Settings) to instantly open Karma and jump to the Redactor page.

HTTP API

Enable the built-in HTTP server in the Server page to expose the redaction API:

# Health check
curl http://127.0.0.1:8000/health

# Redact text
curl -X POST http://127.0.0.1:8000/redact \
  -H "Content-Type: application/json" \
  -d '{"text": "My name is John and my email is john@example.com"}'

Build from Source

Prerequisites

• Node.js >= 20
• Rust (stable)
• CMake and a C++ toolchain (builds the mlx-c wrapper on macOS)

Steps

# Install frontend dependencies
npm install

# Build Tauri app (inference engine is compiled into the app; no sidecar needed)
make app

Or run the full pipeline:

make all

Note: on macOS the build downloads the prebuilt MLX runtime (libmlx.dylib + mlx.metallib, ~130 MB) from the mlx_metal PyPI wheel into src-tauri/vendor/mlx-dist/ via make fetch-mlx (also triggered automatically by make app/make dev). Set PIP_INDEX_URL to a PyPI mirror if pypi.org is unreachable. Remove it with make distclean.

Acknowledgements

• Handy