Pick a folder
Any project on your disk. Nook works on a private copy, so the agent can try, fail and try again without touching your files.
- Nook
- Coding agent
A local AI coding agent. Your code never leaves the machine.
Point Nook at a folder and say what you want changed. A coding model on your GPU reads the project, edits the files and runs your tests on a private copy, then hands you the diff to keep or throw away.
Windows 11 · NVIDIA 8 GB or more recommended · No account, no cloud, no upload
checkout · CodeDone · 3 files changed
YouAdd a discount code field to the cart, with tests.
14 tests passed · Qwen3-Coder, on your GPU · nothing sent anywhere
How it works
Three steps, and your folder is only touched when you say so.
1
2
Say what you want
In plain words: add a field, fix a failing test, rename a class across the project. The model reads the code it needs and edits the files.
3
Review the diff
Nook runs your tests and hands you the changes. Keep what you like, throw away the rest.
Why run the coding agent locally
Cloud coding agents send your repository to someone else's servers on every request. A local one doesn't.
Your code stays yours
Client work under NDA, proprietary code, a codebase you are not allowed to upload: the model runs on your card and nothing is sent. Everything Nook serves listens on 127.0.0.1 only.
No rate limit, no bill per token
Open models answer as often as you ask. No usage cap that resets at the end of the week, no metered API key.
It works offline
Once a model is downloaded, the agent works on a train, on a plane, or with the network cable pulled.
You choose the model
Qwen3-Coder and gpt-oss to start, or any GGUF on Hugging Face, sized against your card before you download it.
Works with the agents you already use
Nook is also an MCP server. Claude Code, Claude Desktop, Codex and Cursor can hand it the work that should stay on your machine: search a repository by meaning, pull the lines of a long log that answer a question, transcribe a recording, make a picture. One click on Home writes the entry into each tool's configuration, with a backup beside the file. What the agent then does with a result is up to the agent and its provider.
What you need
| Requirement | |
|---|---|
| System | Windows 11. Per-user install, no admin rights, no WSL, no account. |
| Graphics card | An NVIDIA card with 8 GB runs it; more memory runs bigger coding models. AMD and Intel cards run through Vulkan. |
| Engine | llama.cpp, placed on your GPU by the Nook Runtime, which sizes each model to your VRAM. |
| Models | Qwen3-Coder, gpt-oss, or any GGUF from Hugging Face. |
Questions
Does my code leave my machine?
No. The coding model runs on your GPU, the agent works on a private copy of your folder, and nothing you type or change is sent anywhere. The only downloads are the models themselves, from Hugging Face.
Which models does the coding agent use?
Qwen3-Coder and gpt-oss to start. You can also browse Hugging Face for any GGUF model; Nook checks it against your card's memory before you download it.
How does it compare with cloud coding agents?
Open models have closed much of the gap. For a change in a codebase they are more than enough, and they answer with no rate limit and no bill. The very largest cloud models still lead on the hardest reasoning; Nook lets you run the biggest model your card can hold.
Can Claude Code, Codex or Cursor use it?
Yes. Nook is an MCP server over Streamable HTTP at 127.0.0.1:41434/mcp, or Nook.exe --mcp over stdio. Connect a tool with one click on Home.
Can I run a local coding model on 8 GB of VRAM?
Yes. An NVIDIA card with 8 GB runs Nook, and more memory lets you run a larger coding model. AMD and Intel cards work through Vulkan.
Everything else Nook does
One app for the AI you run at home: a coding agent, pictures, video, and speech to text with Whisper. All of it on your GPU.
Own your AI.
A Windows desktop app: one download, no account, no admin rights, and it keeps itself up to date.