I've been looking for a local AI agent that doesn't require a terminal, a config file, or a weekend of setup. Local Code is the closest I've found. It's a free, open-source AI agent for Mac by the Draw Things team, and it went fully open source on October 28, 2026 after a public beta that started September 12. No account, no API key, no separate inference server. Download a model from the built-in picker and start working.
The thing that got me to try it was the promise: everything built in. Python 3.14, Node.js 24, a shell, Git, document tools, a browser the assistant can control, and access to Mail, Messages, Notes, Calendar, and Photos. All without installing anything separately. That's a bold claim for a free app.
What Local Code actually is
Local Code is a desktop app that runs Qwen or DeepSeek on your Mac. You pick a model from the built-in selector, it downloads, and you start working. The app processes prompts and files on your Mac rather than sending them to a cloud provider. With models set to run on device, prompts, files, and tool output stay on your Mac, and the app keeps working with Wi-Fi off.
The agent runs inside Apple's App Sandbox and Hardened Runtime, so it can only touch folders you share and can't launch other apps. That's a real constraint, but it's also a real security feature. You're not giving an AI agent root access to your machine. You're giving it access to a workspace folder and the ability to read the Apple apps you explicitly grant.
The thing that surprised me
The built-in tools. Most AI agents are a chat interface that can run shell commands. Local Code is a chat interface that can run shell commands, write Python, write JavaScript, generate images and videos through the Draw Things CLI, create presentations, edit Word and PowerPoint files, analyze spreadsheets, and browse the web. All of it built in. All of it running locally.
The image and video generation is the part I didn't expect. The Draw Things team built Local Code on top of their image generation engine, so the agent can generate images, render motion graphics to video, and create infographics with precise charts. That's not something you get from a terminal agent. It's a desktop app that happens to have a coding agent inside it, not a coding agent that happens to have a desktop app.
What it gets right
The privacy story is the strongest feature. Everything runs on your machine. No cloud, no API keys, no data leaving your Mac. For anyone working with proprietary code or sensitive documents, that's not a nice-to-have, it's a requirement. Local Code makes it the default, not an option you have to configure.
The model picker is well done. You select a model, it downloads, and you start. No terminal, no Ollama setup, no config files. For someone who has never run a local model before, this is the lowest friction path I've seen. The app handles the inference server, the model management, and the agent runtime. You just talk to it.
The offline capability matters more than I expected. I used it on a flight with no Wi-Fi and it worked perfectly. The model was already downloaded, and every tool ran locally. That's the kind of thing you don't appreciate until you need it.
Where it still frays
The hardware requirements are steep. Qwen 3.8 27B needs 16GB of RAM minimum, with recommended downloads taking 10GB to 21GB of storage. The larger DeepSeek models need 89GB or 376GB. That's not a casual download. If you don't have a Mac with enough memory and storage, Local Code is not for you. An M-series Mac with 16GB is the floor, and 32GB is where it gets comfortable.
The Mac-only limitation is real. There's no Windows or Linux version. The developers say a Mac App Store release is coming soon, but for now it's a direct download from getlocalcode.app. If you're not on a Mac, this tool doesn't exist for you.
The model quality ceiling is the same as every local agent. Qwen 3.8 27B is capable but not frontier. For straightforward coding tasks, refactors, and boilerplate, it's good enough. For complex multi-file reasoning, it falls short of Opus 5.5 or GPT-6. The tool is the harness. The model is the engine. Don't blame the harness for the engine.
The App Sandbox is both a feature and a limitation. It keeps the agent contained, which is good for security. But it also means the agent can't do everything a terminal agent can. If you need to install system packages, modify system configs, or run arbitrary shell commands outside the sandbox, you'll hit walls. For most coding tasks, the sandbox is fine. For advanced workflows, it's a constraint.
The honest comparison
I ran Local Code alongside OpenCode for a week. Here's what I found.
For local-only work, Local Code wins outright. No terminal, no config, no Ollama setup. Download a model and go. OpenCode is more flexible, but flexibility costs setup time. Local Code trades flexibility for convenience, and for most users that's the right trade.
For model variety, OpenCode wins. It talks to 20+ providers. Local Code runs Qwen and DeepSeek locally, plus cloud models through your own API keys. If you want to use a specific model, OpenCode gives you more options.
For tool breadth, Local Code wins. Python, Node, shell, Git, document editing, image generation, web browsing, all built in. OpenCode has a plugin system, but the built-in toolset is narrower. If you want an agent that can do more than code, Local Code has the broader toolkit.
For privacy, Local Code wins. Everything runs on device by default. OpenCode can run locally too, but the setup is more involved. Local Code makes privacy the default, not an option.
Who should use it
If you're a Mac user who wants a local AI agent without touching a terminal, Local Code is the best option I've found. It's not perfect, but it's the lowest-friction path to running a capable model on your own machine.
If you're a developer who already uses OpenCode or Goose and you're happy with them, Local Code won't convert you. The terminal agents are more flexible and more powerful. But if you've been avoiding local models because the setup is intimidating, Local Code removes that barrier.
If you're on Windows or Linux, this tool doesn't exist for you. That's the biggest limitation, and it's the reason I can't recommend it universally.
The honest verdict
Local Code is the most accessible local AI agent I've tried. That's a meaningful distinction, because accessibility is the thing that keeps most developers from running local models. The hardware requirements are steep, the platform is Mac-only, and the model quality ceiling is real. But for a Mac user with enough RAM who wants privacy and convenience, it's the best option available.
It's not as flexible as OpenCode. It's not as polished as Cursor. But it will not charge you, it will not send your code to the cloud, and it will work on an airplane. For a lot of developers, that's the deal they've been waiting for.
Would you trade flexibility for convenience, or does the terminal still win for you?
Top comments (0)