Well, the very next sentence after your quote is "It requires you to supply an OpenAI API key" which should answer your question.

Here's the default prompt:

> Return commands suitable for copy/pasting into \(shell) on \(uname). Do NOT include commentary NOR Markdown triple-backtick code blocks as your whole response will be copied into my terminal automatically.

> The script should do this: \(ai.prompt)

This doesn’t answer my question about watching the terminal contents. If I ‘cat’ a file, is the file’s contents sent to openai?

When you open Codecierge in the toolbar sidebar, it warns you with this:

> Everything that happens in your terminal while Codecierge is running will be sent to OpenAI. Don't send them confidential information!

And when you check the box "Run commands automatically" it shows you this popup:

> This lets an AI completely control your computer. It could delete your files, do something stupid or dangerous, or lead to the downfall of humanity. Proceed with caution.

Though that's the Codecierge feature "Toolbelt -> Show Toolbelt" + Toggle "Codecierge" toolbelt window. "Run commands automatically" is hilarious.

The simpler Edit -> Engage AI (Cmd+Y) feature seems much more useful for day to day stuff and it doesn't send terminal contents to OpenAI. e.g. Cmd+Y -> "Extract foo.tar in verbose mode".

That renders the LLM's response command in a small pane and you have to Shift+Enter to go ahead and paste it into the terminal. That's really nice.

> "Run commands automatically" is hilarious.

Yeah they should make it so it mocks the user for attempting to set it to true and explain in select, direct, simple to understand words why that's a dumb idea and that obviously the user didn't read the disclaimer.

Yeah, it kinda seems like a feature a developer would implement for fun but I don't see much practical use for it off the top of my head.

In the sidebar Codecierge convo they do give you a "Copy code" button whenever it detects code blocks. I think it'd be more practical to also add a "Paste and run" shortcut to each code block. i.e. A little "Play" button.

Aside, Codecierge doesn't seem to read the terminal contents like it says it should. If you ask it any questions or reference anything in your terminal, like an error in response to a command it gave you, it doesn't seem to have that context.

> Don't send them confidential information!

Most of what I do in a terminal is confidential. I find it scary and also funny that in the age of data laws, privacy policies, data leaks, etc, someone thought this is an acceptable workflow.

For someone working on their personal projects for which they can make decisions about visibility, why wouldn't it be an acceptable workflow?

Absolutely fine for that workflow.

Make sure to add a note to the project’s privacy policy that user data is sent to openai.

You can set a custom URL to use Ollama which is OpenAI API compatible. Llama3 8b runs quite fast for me on a M1 Max. I'll be making use of this feature I think (haven't tried it yet).

No, unfortunately it's not possible to set a custom URL... OpenAI is either hard coded or using a configuration option that isn't exposed.

It is exposed from Advanced tab under "URL for AI API". You do have to change the model name from the preselected and offered OpenAI GPT model names. Otherwise this AI API URL isn't used. Haven't yet tested with local OAI compatible model yet, though, but did manage to break it using bad address :)

How do you change the model name from the preselected and offered OpenAI GPT model names? I don't see the option to do so.

click the name instead of the dropdown arrow

Note that to make use of ollama, you'll have to wait for the next release or upgrade to a beta release --- while you can change the API right now it's using a newer format for the payload, that's fixed in the beta/next release.

Advanced settings -> "URL for AI API"

If this is a local pattern matcher trained on some text pool it could be really useful.