aicommit is a small command line tool for generating commit messages. There are many of these already out there, some even with the same name. But none (to my knowledge) follow the repository's existing style, making them useless when working in an established codebase.

A good commit message is more than a summary of the code changes. It contains the intention, context, and external references that help others understand the change. Thus, aicommit has a -c/--context flag for quickly adding this detail.

aicommit is inspired by our good friend @sreya:

Via Homebrew:

brew install aicommit

Or, using Go:

go install github.com/coder/aicommit/cmd/aicommit@main

Or, download a binary from the latest release.

You can run aicommit with no arguments to generate a commit message for the staged changes.

export OPENAI_API_KEY="..."
aicommit

You can "retry" a commit message by using the -a/--amend flag.

aicommit -a

You can dry-run with -d/--dry to see the ideal message without committing.

aicommit -d

Or, you can point to a specific ref:

aicommit HEAD~3

You can also provide context to the AI to help it generate a better commit message:

aicommit -c "closes #123"

aicommit -c "improved HTTP performance by 50%"

aicommit -c "bad code but need for urgent customer fix"

When tired of setting environment variables, you can save your key to disk:

export OPENAI_API_KEY="..."
aicommit --save-key
# The environment variable will override the saved key.

aicommit will read the COMMITS.md file in the root of the repository to determine the style guide. It is optional, but if it exists, it will be followed even if the rules there diverge from the norm.

If there is no repo style guide, aicommit will look for a user style guide in ~/COMMITS.md.

You may set OPENAI_BASE_URL to use other OpenAI compatible APIs with aicommit. So far, I've tested it with LiteLLM across local models (via ollama) and Anthropic. I have yet to find a local model that is well-steered by the prompt design here, but the Anthropic Claude 3.5 commit messages are on par with 4o. My theory for why local models don't work well is they (incl. "Instruct" models) have much worse instruction fine-tuning than flagship commercial models.