cai is a Git extension that automates the creation of commit messages.
Simply run git cai to generate a meaningful, context-aware commit message based on the changes in your repository.
cai uses a Large Language Model (LLM) to analyse diffs and new files, producing concise and informative commit messages.
Currently supported providers:
- OpenAI
- Gemini
- Anthropic
- Groq
- xAI
- Mistral
- DeepSeek
- Ollama (local)
- Python 3.10 or higher
- pipx
- Either:
- Ollama installed and running locally, or
- An API key for at least one of the following providers:
- OpenAI
- Gemini (free tier available)
- Anthropic
- Groq (free tier available)
- xAI
- Mistral
- DeepSeek
- Automatically detects added, modified, and deleted files
- Generates meaningful, context-aware commit messages using an LLM
- Seamless integration with Git
- Supports multiple LLM providers and models
- Override provider and model per invocation (
-P,-m) - Global configuration with per-repository overrides
- Repository-specific language, style, and model selection
- Amend the last commit message with a regenerated one (
-A) - Conventional Commits format support (
-C) - Change configuration from the command line (
-S,-H) - Optional commit squashing with automatic summary generation (all, last N, or up to a specific commit)
- List providers, models, active config, and file paths (
-l) - Token usage logging for API calls
- Branch name as LLM context via
--branch/-b - Extra context for the LLM via
--context/-x(e.g. ticket numbers, reasons) - Generation time measurement (
-t) - Shell completion for bash, zsh, and fish (
-i)
Install cai using pipx:
pipx install git-cai-cliEnsure that pipx binaries are available in your PATH:
pipx ensurepathRestart your shell after installation.
If you are running Arch Linux or an Arch-based distribution such as EndeavourOS, CachyOS, etc., you can install the package from the AUR using a package manager like Paru.
paru -S caiOnce installed, cai works like a standard Git command:
git caicai uses the output of git diff to generate a commit message.
The generated message is opened in your configured Git editor, allowing you to review or edit it before committing.
In short: it behaves like git commit, but the commit message is pre-filled.
To exclude specific files or directories from being considered when generating commit messages, create a .caiignore file in the root of your repository.
- Files listed in
.gitignoreare always excluded .caiignoreis intended for tracked files that should not influence commit messages
The syntax is identical to .gitignore.
On first execution, cai automatically creates the base configuration in your home directory.
-
Global configuration:
~/.config/cai/cai_config.yml -
API tokens:
~/.config/cai/tokens.yml
It also creates two Markdown prompt files:
- Default commit prompt:
~/.config/cai/commit_prompt.md - Default squash prompt:
~/.config/cai/squash_prompt.md
Don't be scared the first run will show an error. It only misses a token.
Add your provider API keys to tokens.yml. Once configured, cai will reuse them automatically.
Set your preferred LLM in cai_config.yml (Groq by default).
If you want to use Ollama, set default: ollama and configure the ollama: block (model/temperature). Ollama is automatically started when used.
The generated commit message is guided by prompt files.
- By default,
cai_config.ymlpoints to the auto-created prompt files in~/.config/cai/. - To use your own prompts in a repository, generate templates at the root of the repository:
git cai -pThis creates:
commit_prompt.mdsquash_prompt.md
Then set prompt_file and/or squash_prompt_file in your cai_config.yml (also repo) to point to those files.
Each repository can be configured independently.
If a cai_config.yml file exists in the root of a repository, cai will use it instead of the global configuration.
This allows different projects to use different providers, models, languages, and styles.
Examples of per-repository customization:
- Different LLM providers or models
- Different commit message languages
- Different writing styles or tones
- Emojis enabled or disabled per project
To create a repository-specific configuration:
cp ~/.config/cai/cai_config.yml .Modify the copied file as needed. As an alternative execute:
git cai -gdefault– default LLM providermodel– model to use for the selected provider
(note: not all models may be compatible)temperature– controls how creative the generated messages arelanguage– language used for commit messagesstyle– tone or style of the commit messageemoji– enable or disable emojisload_tokens_from– path to the file where API tokens are storedprompt_file- path to the file where the prompt for the commit is storedsquash_prompt_file- path to the file where the prompt for the squash is storedbranch_context– include current branch name as LLM context (default:false)conventional– use Conventional Commits format (default:false)token_logging– log token usage after each LLM call (default:truefor new installs)measure_time– log generation time (default:false)
In addition to git cai, the following options are available:
-A,--amend– regenerate and amend the last commit message-a,--all– stage all tracked modified and deleted files-b,--branch– include current branch name as context for the LLM-C,--conventional– use Conventional Commits format (type(scope): description)-c,--crazy– Trust the LLM and commit without checking-d,--debug– enable debug logging-g,--generate-config– generate the defaultcai_config.ymlin the current directory-H,--set-home– set a config value in home config (key=value), always targets~/.config/cai/-h,--help– show help and available commands-i,--install-completion– install shell completion for bash, zsh, or fish-l,--list– list available information (config,editor,language,model,path,provider,style)-m,--model– override the model for this invocation (requires-P)-p,--generate-prompts– generate defaultcommit_prompt.mdandsquash_prompt.mdin the current directory (for customization)-P,--provider– override the LLM provider for this invocation-S,--set– set a config value (key=value) in repo config (requires existing repo config)-H,--set-home– set a config value in home config (key=value), always targets~/.config/cai/-s,--squash[N|HASH]– squash commits on the current branch and summarize them. Without argument: squash all since branch checkout. With a number: squash the last N commits. With a commit hash: squash up to and including that commit-t,--time– measure and log commit message generation time-x,--context– provide extra context for the LLM (e.g. ticket number, reason for change)-u,--update– check for updates-v,--version– show the installed version
To regenerate the last commit message and amend it:
git cai -AThis reads the diff from the most recent commit, sends it to the LLM, and opens the editor for review.
Use with -c to amend immediately without the editor: git cai -A -c.
To generate commit messages in Conventional Commits format:
git cai -CThis enforces the type(scope): description structure. Allowed types: feat, fix, docs, style, refactor, perf, test, build, ci, chore, revert. Use ! after the type/scope for breaking changes.
To enable it permanently:
git cai -S conventional=trueInstead of editing YAML files manually, use --set or --set-home to update config values.
--set (-S) targets the repo config (requires an existing cai_config.yml in the repo root):
git cai -S default=anthropic # change the default provider
git cai -S emoji=false # disable emojis
git cai -S groq.model=llama-3.3-70b # nested key (dot notation)
git cai -S openai.temperature=0.7 # set temperature as floatIf no repo config exists, an error is shown. Use git cai -g to create one first.
--set-home (-H) always targets the home (default) config (~/.config/cai/):
git cai -H language=de
git cai -H emoji=falseThis project is licensed under the MIT License.