# Local Code

> A private, open-source AI agent for Mac, iPad and iPhone that runs models on-device with built-in Python, Node.js and a shell.

Local Code is a private AI agent made by Draw Things. It combines an LLM inference engine for Apple Silicon, a coding harness tuned to each model, and an agent runtime with Python, Node.js and a shell. You give it a workspace folder and a concrete task, then review the files it produces.

## What It Is

Local Code is an on-device agent app for Mac, iPad and iPhone. It works on files you choose, such as analyzing data, charting it, writing documents and slide decks, or building software. With an on-device model, inference runs locally and works offline once the model is downloaded. Cloud models from Anthropic, OpenAI, OpenRouter or an OpenAI-compatible server are optional, and in that case prompts, file contents and tool output go to the chosen provider.

## Sandbox and tools

The agent's shell, language runtimes and command-line tools are compiled into the app and run inside Apple's App Sandbox, with Hardened Runtime on Mac. The agent can reach only the folders you allow and cannot launch other programs. The sandbox includes an OSH shell with common Unix commands, Python 3.14 with NumPy, pandas and Matplotlib, Node.js 24, Lean 4 and Z3, LibreOffice for Office and PDF files, FFmpeg, web search, a drivable WebKit browser, and the Draw Things CLI for image and video generation. Calendar and photos commands work after access is granted. Built-in skills cover Office files, charts, HTML-to-video, static site deployment, MCP servers via MCPorter and skill creation. Users can add their own skills.

## Models and memory

Local Code downloads its own quantized checkpoints of DeepSeek 4.1 Flash 1017, DeepSeek 4 Flash 0731 and Qwen 3.8 27B. The DeepSeek models are mixtures of experts that stream part of their weights from the SSD, so they can run on Macs with less memory than the checkpoint size. Setup suggests a model based on device memory. The project publishes its own speed, benchmark and quantization comparisons in the repository README.

## Availability

On Mac it needs Apple Silicon with 16 GB or more and macOS 26.0 or later. It is available on the App Store and as a direct Mac download. The source is released under GPL-3.0.

## Features
- On-device LLM inference on Apple Silicon
- Works offline once a model is downloaded
- Sandboxed agent runtime with Python, Node.js and shell
- Built-in skills for Office files, charts, video, images and static sites
- Durable tasks and threads
- Resumable generation
- Steering the agent mid-task
- Long-term context cache on disk
- SSD weights streaming for large models
- Optional cloud models
- Custom skills
- MCP server support via MCPorter
- Open source under GPL-3.0

## Integrations
DeepSeek, Qwen, Anthropic, OpenAI, OpenRouter, OpenAI-compatible servers, Draw Things CLI, LibreOffice, FFmpeg, Cloudflare Wrangler, Model Context Protocol

## Platforms
MACOS, IOS, WEB, API, CLI

## Pricing
Open Source, Free tier available

## Links
- Website: https://getlocalcode.app
- Documentation: https://getlocalcode.app/guides
- Repository: https://github.com/drawthingsai/local-code-community
- EveryDev.ai: https://www.everydev.ai/tools/local-code