One zip. No installer.
CodAI v3.0.0 ships as a portable Windows build: the official llama-server engine and a browser chat UI in one folder, with a controller that keeps the engine healthy. Unzip, drop in a GGUF model, double-click run.bat.
v3.0.0 · .zip · Windows 10/11 (64-bit). Verify the release on GitHub before running.
System requirements
- Windows 10 / 11 (64-bit)
- 4 GB RAM minimum — 2 GB works with the smallest models
- ~1 GB free for the app + 0.5–5 GB for one model
- No GPU required — the engine runs on CPU
macOS & Linux
v3.0.0 ships Windows-first. On macOS or Linux you can run the same engine yourself — llama.cpp builds on all three — or compile the controller from source. Older cross-platform releases remain on GitHub.
Build from source →What’s in the box.
Codai.exe
The packaged controller — starts the engine, monitors health, auto-restarts after crashes.
engine\llama-server.exe
The official llama.cpp build. Your prompts never touch anything but this process.
run.bat
Starts everything, waits for [READY], opens the chat UI in your browser.
kill.bat
Force-cleans Codai processes and locks when you want everything stopped.
First run, three steps.
- 01
Unzip anywhere writable.
Not C:\Program Files — the folder holds the engine, your models, and rotating logs, so pick a place it can write to. A USB stick works.
- 02
Drop one model into models\.
Any GGUF works. The default pick is gemma-3-1b-it-Q4_K_M.gguf (0.81 GB) — small enough for 2 GB-RAM machines.
- 03
Double-click run.bat.
Wait for [READY] in the console. The chat UI opens in your browser at a local address, streaming tokens from your own hardware.
Air-gapped by design.
Air-gapped operation
Fully functional without any active internet connection or network interface.
Zero code telemetry
Keystrokes, source code, and embeddings never leave local RAM.
Local model weights
GGUF quantized weights run locally via the official llama.cpp engine.
GPL-3.0 license
The controller’s license terms are published in the repository — read them before or after you download.
Or compile it yourself.
Audit the codebase, then build CodAI from source. If you don’t trust the zip, this is the path that removes the question entirely — and it’s also the macOS/Linux route.