Access the most powerful LLMs from within the terminal.

• Streamed responses with full Markdown rendering and syntax highlighting
• Intentionally minimal UI/UX with configurable styling
• Support for OpenAI and Anthropic models, with more coming soon (Mixtral, DBRX)
  • Easily choose model with CLI option and API key
  • Highly extensible implementation
• Top-class terminal integration via escape codes

I created this out of curiosity of CLI/TUI development and because of the lack of free and convenient interfaces to GPT-4. Now a staple in my daily workflow as a rubber duck debugger and google alternative, I have found this CLI works great within one's IDE.

Developed on MacOS and Ubuntu. All instructions will assume a unix-like OS with Python 3.11+ installed.

An OpenAI API key is required to access the GPT models and an Anthropic API key is needed for the Claude models. Both of these companies currently require paying for usage credits in advance, so ensure you have an account with money loaded before using the CLI.

Sign up for OpenAI here and Anthropic here.

Once you have you API keys, place them in a file named env.json in the project root directory like so:

gpt-cli/env.json

{
  "openaiAPIKey": "12345678",
   "anthropicAPIKey": "87654321"
}

I prefer to install this app's dependencies to the system interpreter so that I can use a shell alias to run it at any time with no prior venv set up. In the terminal:

1. Ensure you are not using a venv: run deactivate
2. After cloning and inside the project directory, run make install to download all python packages.
3. Add shell alias and reload shell config to apply changes (assuming bash or zsh):

   config_file="${HOME}/.$(basename ${SHELL})rc" && \
   echo -e "\nalias llm='python3 ENTER_ABSOLUTE_PATH_TO_/gpt-cli/src/gpt-cli/main.py_HERE'" >> "${config_file}" && \ 
   source "${config_file}" 

Now no matter where your current working directory is, you can type llm in your terminal to start a GPT session!

Note: All settings are documented in the help page, accessible by running with the help option: llm --help (assuming you have the shell alias described above)

This is extremely important to properly tailor to your use case. Heavily influences every response from the model. By default, it is "You are a concise assistant to a software engineer", but you can change this at any time via command line arguments or by editing the source code.

styling.py contains many editable values to control the look of the program

Markdown code blocks can be rendered in any of the styles found here by setting the --code-theme option, and the color of the plaintext responses can be set to most normal colors using --text-color (ex. llm --text-color=orange)

The program is a simple REPL. Each time you click Enter your prompt will be sent and the response will be streamed in real time. CTRL+D and CTRL+C work as expected in a REPL, exiting the program and cancelling the current loop, respectively. Entering q, quit, e, or exit while in prompt mode will also exit the program.

When typing a prompt, basic keyboard shortcuts are available like history navigation with the arrow-keys and deleting entire line with CTRL + U. More will be added in the future. Multiline text input is supported by entering ml a single backslash into the prompt. This is useful when pasting code into the prompt.

When in multiline mode, you must use Meta + Enter to submit the prompt. On macOS it is [Option or Command] + Escape + Enter

Your entire chat history will be sent to the language model each time you press Enter. To clear your chat history, enter c or clear into the prompt. I recommend doing this whenever you want to ask the model a different line of questioning, so that you get the highest quality answers and you do not run up your usage bill.

Once your current conversation costs more than a cent or two, it will be shown at the end of the response so that you know how much you're spending. Total session cost will also be shown when the program exits.

Disclaimer: While I do use the official OpenAI tokenizer and an official cookbook example from that repo to calculate prices, it is not something I have tested thoroughly, so there is no guarantee the prices are accurate. Anthropic provides authoritative token counts however, so those should be calculated better

The pricing for the GPT turbo models is pretty cheap. Other than the Opus model, Anthropic's offerings are comparatively even cheaper. For clarity, both OpenAI and Anthropic bill you on tokens per request AND response, and the entire chat history of previous prompts and responses are sent in EACH request. As long as you remember to clear your history after a few prompts you will be fine.

If you want to have multiple sessions, use screen or tmux. I want this project to "do one thing well" and leave other features to existing programs. This is the difference between this project and elia, for example.

This is a personal project above all else. I use this to play around with design patterns, new python features, data structures, best practices, etc.

Feature requests are more than welcome, however I will probably take a while to get to them.

• Ensure the code-base is highly extensible
• Support for other LLMs if any seem worth adding
• Incorporate cutting-edge tooling like uv and Ruff.