Álvaro Ramírez

UPDATE: chatgpt-shell has evolved a bit and is now on MELPA.

I had been meaning to give ChatGPT a good try, preferably from Emacs. As an eshell fan, ChatGPT seemed like the perfect fit for a shell interface of sorts. With that in mind, I set out to wire ChatGPT with Emacs's general command interpreter (comint).

I had no previous experience building anything comint-related, so I figured I could just take a peek at an existing comint-derived mode to achieve a similar purpose. inferior-emacs-lisp-mode (ielm) seemed to fit the bill just fine, so I borrowed quite a bit to assemble a basic shell experience.

From then on, it was mostly about sending each request over to the ChatGPT API to get a response. For now, I'm relying on curl to make each request. The invocation is fairly straightforward:

curl "https://api.openai.com/v1/chat/completions" \
     -H "Authorization: Bearer YOUR_OPENAI_KEY" \
     -H "Content-Type: application/json" \
     -d "{
     \"model\": \"gpt-3.5-turbo\",
     \"messages\": [{\"role\": \"user\", \"content\": \"YOUR PROMPT\"}]
     }"

There are two bits of information needed in each request. The API key, which you must get from OpenAI, and the prompt text itself (i.e. whatever you want ChatGPT to help you with). The results are not too shabby.

I've uploaded the code to GitHub as a tiny chatgpt-shell package. It's a little experimental and rough still, but hey, it does the job for now. Head over to github to take a look. The latest iteration handles multiline prompts (use C-j for newlines) and basic code highlighting.

Let's see where it all goes. Pull requests for improvements totally welcome ;-)