Offhand
Free · Lightweight commit model · Your data stays local and private.
Stage your changes in VS Code's terminal and get a commit message suggested right there:
$ git add .
Offhand suggests: feat: add subtract function
Commit with this message? [Y/N] y
✓ Committed 3f2a1bc feat: add subtract function
Y commits with the suggestion. N lets you type your own message (empty cancels). The messages
follow Conventional Commits (feat:, fix(scope):, docs: ...).
It runs a small fine-tuned model on your machine: no account, no API key, and your code never leaves
your computer.
Getting started
- Click Offhand: Off in the bottom-left status bar to turn it On. The first time, it
downloads the engine and the model (about 1 GB, once); progress shows in the status bar.
- Open a terminal in VS Code, save your files, and run
git add ..
- Answer Y or N.
Click Offhand: On to turn it off and free the memory it uses (about 1 GB).
Only git add in VS Code's terminal triggers it: other git commands, the + buttons in Source Control,
scripts, and everything while it's off work exactly as before. Terminals that were open before you
installed the extension need to be reopened once.
The ✨ button in the Source Control panel puts a suggestion in the commit box instead.
Requirements
- Git, and a VS Code terminal: bash, zsh, Git Bash, PowerShell or Command Prompt.
- About 1.1 GB of disk space for the model and engine, and about 1 GB of memory while it's on.
- A GPU is used when available (NVIDIA, AMD and Intel on Windows and Linux, Apple Silicon on macOS);
otherwise it runs on the CPU (a second or two per message).
- Windows: the engine needs the Microsoft Visual C++ runtime. Many PCs already have it; if yours
doesn't, the extension offers to install it from Microsoft (click Install it; Windows asks for
permission).
Privacy
Everything runs locally. The only network access is the one-time download of the engine (from
llama.cpp's GitHub releases) and the model (from Hugging Face), and on Windows, if you accept, the
Microsoft runtime installer.
Good to know
- Only the first line (subject) is written. Read it before typing Y.
- The model tends to label too many changes as
fix, and rarely picks build, ci, perf or style.
Its descriptions are usually accurate even when the type is off.
- This is a preview release. It has been tested on Linux and Windows 11; macOS should work but hasn't
been tested yet.
Commands
Offhand: Turn On / Off, Turn On, Turn Off, Suggest Commit Message in the Commit Box
(Command Palette, Ctrl+Shift+P). Activity log: Output panel → Offhand.
Settings
| Setting |
Default |
|
offhand.gpu |
auto |
cpu to always run on the CPU |
offhand.modelUri |
hf:Jess2005/commit-model-CLI/commit-model-Q4_K_M.gguf |
Where the model is downloaded from |
offhand.modelPath |
empty |
A local .gguf model to use instead of downloading |
The model
Jess2005/commit-model-CLI: Qwen2.5-Coder-1.5B-Instruct
fine-tuned with LoRA on CommitBench, quantized to 4 bits (986 MB). On 9,704 held-out real commits it
writes a valid Conventional Commits message 99.9% of the time, with the right type 67% of the time.
Because of CommitBench's license, the model is for non-commercial use. See NOTICE.md.
License
The extension's code is MIT licensed. The model and the engine are downloaded on first start and keep
their own licenses, listed in NOTICE.md.
Development
Source and the training pipeline: https://github.com/AhmedJesserLahmer/commit-model
cd vscode-extension
npm install && npm run compile
npm run test:unit # suggestion cleanup
npm run test:integration # the whole workflow inside a real VS Code
Press F5 in the vscode-extension folder to run it in a development window. Diff filtering and prompt
building in src/prompt.ts mirror src/commit_model/diff_utils.py: keep the two in sync.
