AI settings

AI commit messages are configured in Innesta ▸ Settings… (⌘,) → AI, and all of it is optional. Innesta works with the whole feature switched off.

One repository can differ from the rest. Repository AI, in that repository’s own settings, overrides the provider, the model or the style for it alone, and leaves the others as they were.

The provider

Apple Intelligence, on your Mac, when your Mac supports it. Nothing leaves the machine.

Claude Code CLI or OpenAI Codex CLI, in the Direct build, and only after you have explicitly agreed. These are command line tools you have installed and authenticated yourself; Innesta runs one of them locally, and what it does with the diff afterwards is between you and that provider.

Finding the CLI

Innesta looks for the tool on your PATH. If it is somewhere unusual you can give it the path directly.

Models

Codex has a model picker, with the available models refreshed from the tool itself, so the list is what your account can actually use rather than a list compiled when Innesta was built. Claude Code’s model is configurable too.

Style

Five styles — imperative, Conventional Commits, gitmoji, descriptive, and repo-match. See AI commit messages.

Authentication variables

Innesta can forward the OpenAI and Anthropic authentication variables from your environment to the CLI, from an allowlist and only when you have switched it on. It forwards the named variables and not the rest of your environment.

What is done to the request

• A timeout, so a CLI that hangs does not hang the app.
• The output is sanitised before it reaches the message field.
• The staged diff is passed with prompt-injection mitigations — a repository can contain text written to be read by a model, and a commit message drafted from it should not do what that text says.