AI-powered command generation from natural language (Cmd+Shift+. overlay)

[Nurozen]

Summary

A hotkey-triggered overlay that lets users describe what they want to do in natural language and generates the corresponding shell command, which is then inserted into the terminal. Similar to Cursor's Cmd+K in terminal or Warp's AI features.

An AI assisted version is forked here:
https://github.com/Nurozen/ghostty/tree/macos-ai-command

Motivation

Remembering exact flags and syntax for infrequently-used commands (tar, find, ffmpeg, rsync, etc.) breaks flow. Being able to type "find all .log files older than 30 days and delete them" and get the correct command directly in the terminal would be useful, especially since the terminal already has the context of where the command will run.

Proposed Behavior

1. User presses keybind (e.g. Cmd+Shift+.)
2. A small overlay appears with a text field - focus is shifted to entry field
3. User types a natural language description and presses Enter
4. The overlay sends the prompt to a configured AI provider
5. The generated command is inserted into the terminal as text (not executed)
6. User can review and press Enter to run it, or edit it first

The command is not auto-executed — it's inserted as text so the user always has the final say.

Configuration

Provider: openai, anthropic, ollama, custom

ai-provider = openai
ai-model = gpt-4o-mini

Or use a local model with no API key needed

ai-provider = ollama
ai-model = codellama

API keys would be read from standard environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY). Ollama requires no key since it runs locally.

Optional overrides: ai-endpoint (custom API URL), ai-system-prompt (prepend custom instructions), ai-max-tokens.

Scope

• macOS only for initial implementation

[jcollie]

AI features like this are not going to be part of the core Ghostty terminal. Eventually we'll have a plugin system and users will be free to develop plugins that provide these features.

[Nurozen]

thanks for the info @jcollie

[8bittts]

@Nurozen i'm actually interested in this too! love to hear more about your thoughts. there are orchestrators (via cli) that are pretty good but it seems you're wanting it even more tightly integrated.

[Nurozen]

@8bittts I think this is more of a tradeoff that it used to be, but hasn't necessarily reached a tipping point.

Pre-AI, it made tons of sense to go the plugin/3rd party route since the time/money cost of tightly integrating every feature yourself was extreme. Even as a community project, this held due to things like exposure cap. (e.g., only so many people will ever see a project and a small portion contribute.)
In that world, supporting a set of plugins in a standardized way is extremely beneficial as the ROI of each one is rather low but the plugin system is much higher.

That said, things have certainly changed and it might be that the above benefits hold up, for other reasons like unified experience supporting a minimal tool change over anxiety/learning gap.

I do think that argument has lost some ground post-AI, this is a good thing through as it gives us more overhead for thoughtful decision making about how we want to structure projects.

Whether or not its enough of a loss to change many or most projects is still up in the air at this point imo.

So, as long as there is good reasoning, I'm ok with either approach.

If you're keen, feel free to check out my fork linked in the discussion.