ChatGPT CLI

ChatGPT CLI is a command-line interface tool that interacts with OpenAI's ChatGPT API. It allows users to communicate with GPT models, adjust API parameters, use magic commands, insert file contents, fetch content from URLs, and execute system commands while chatting.

Features

• Interactive command-line interface using GNU Readline with Emacs keyboard shortcuts.
• Expand file contents from the file path(s) using the placeholder.
• Expand web page contents from the URL using the placeholder.
• Magic commands to clear, undo, edit, write, save, and load data.
• Execute system commands and pass captured stdout and stderr output to ChatGPT.
• Code blocks in the response are saved in temp files and can be referenced from $CODE1, $CODE2...
• Output execution results in HTML (experimental).
• Substitution patterns of placeholders can be configurable in the configuration file.

Installation

GtiHub Release

Download binaries from Github Release for Linux or macOS.

Note: Binaries for macOS are not statically compiled and require shared libraries, such as libssl and libcrypto. If you encounter any errors, please check the required libraries by running otool -L /usr/local/bin/chatgpt.

Windows is not supported.

From source code

1. Install Crystal

2. Build the project:

   git clone https://github.com/kojix2/chatgpt-cli
   git submodule update -i # Awesome ChatGPT Prompts
   cd chatgpt-cli
   make
   sudo make install # Or simply copy bin/chatgpt to your directory in $PATH.
   

Set your OpenAI API key as an environment variable by running:

export OPENAI_API_KEY="your_openai_api_key"

Using a proxy server

export HTTP_PROXY=http://[IP]:[port]
export HTTPS_PROXY=https://[IP]:[port]

Usage

To start using ChatGPT CLI, run the chatgpt command in your terminal:

chatgpt

Type a message and press Enter to get a response from GPT-3.5 Turbo. To exit, type exit, quit, or press Ctrl + D.

Options

You can set various options when running the chatgpt command:

Usage: bin/chatgpt [options]
    -r, --resume                     Resume the session
    -m MODEL, --model MODEL          Model name [gpt-3.5-turbo]
    -i ID, --id ID                   ID of the custom system message
    -s STR, --system STR             System message
    -n INT                           Number of edits to generate [1]
    -t Float, --temperature Float    Sampling temperature to use [1.0]
    -p Float, --top_p Float          Probability threshold of nucleus sampling [1.0]
    -v, --version                    Print version info and exit
    -h, --help                       Print help

Restore the previous chat and use the gpt-4 model.

chatgpt -r -m gpt-4

Note that short options cannot be concatenated.

chatgpt -rm gpt-4
# Error: Invalid option: -rm

Using ChatGPT interactively

Start ChatGPT CLI.

chatgpt

Write your message using GNU Readline, which supports Emacs shortcuts, such as ↑ and ↓.

> Hi

You can cancel a query to ChatGPT with Ctrl + C. This is especially useful when you make a writing mistake, and it takes a long time to reply.

Unfortunately, line breaks are not currently supported, but still you can copy-paste line breaks.

Selecting ChatGPT System Commands

Select pre-registered system commands:

chatgpt -i "Linux Terminal"

Awesome ChatGPT Prompts are available.

Passing Files to ChatGPT

You can pass the contents of files to ChatGPT by using the %{foo.txt} pattern.

> Please explain the following code: %{src/beatles.py} Are there any bugs?

With this feature, you can pass the code to ChatGPT and have them write README.md.

> Please read the code of the tool: %{src/*.cr} %{src/**/*.cr} Then update README.md %{README.md}
...
> %w README.md # This will save README.md

Note: READMEs created this way tend to be characterless. Still, it is marvelous to have a computer write the README for you.

%{} can be changed in the config file.

Passing Web Pages to ChatGPT

You can pass the contents of a web page to ChatGPT by using %%{www.example.com}.

> Pick five interesting news items: %%{https://news.ycombinator.com/}

Here, the HTML from the URL is fetched, the words used in the body are extracted and passed to ChatGPT. It is much smaller than the raw HTML, but still not enough.

%%{} can be changed in the config file.

Magic Commands

During the dialogue, you can use a variety of magic commands.

%clear

Clear all messages. This is useful when you want to change the topic and set TOKEN back to zero.

%undo <n>

Undo the last n query and response [1]. If you make a mistake asking ChatGPT or if the reply is not what you are looking for, you can undo.

%shift <n>

Remove first messages and responses. This is useful when you want to continue the chat even after the Token count has reached its limit.

%write <filename>

Write the most recent message to a file. Save the text and scripts written by ChatGPT.

%w <filename>

Alias for write. This magic command is used so often that an alias is provided.

%config

Edit the config file. Open a text editor and rewrite the settings.

%system

Show the current system message.

%system <message>

Set a new system message (this behavior is subject to change).

%edit

Edit data in JSON format. All data passed to ChatGPT can be edited here. You are free to tamper with the past.

%html <filename>

Export the conversation to HTML and launch your browser. This feature is useful when you want to save a conversation or check path expansion, etc. (Experimental and the output HTML may not be correct)

%save <filename>

Save the data. This allows you to do things like "save session". The file is a JSON file to be posted to ChatGPT.

%load <filename>

Load the data. This allows you to do things like "load session".

%resume

Load data from auto-saved data file. We humans forget to save data. Therefore, the last session is automatically saved.

%model <name>

You can change the model used in the next request. For example, if you find your question is too difficult for gpt-3.5-turbo, you can change to gpt-4 at that point.

%tokens

Total tokens used. You will see a more detailed number of TOKENs than is shown in the prompt. Please note that chatgpt-cli does not have the ability to calculate the number of TOKENs. It only displays the ChatGPT response. Therefore, if any edits may have been made, it will be Unknown.

%webapp

Open the ChatGPT website.

%debug

Show debug message. Display the data actually posted to ChatGPT and the response in JSON format.

%help

Show the help. Humans forget commands.

Note that for %config, %data, and other commands launch an editor. The editor used can be set by the EDITOR environment variable. Note that the tool is still being improved and the behavior of the magic commands will continue to change.

Executing System Commands

You can execute system commands while chatting by prefixing the command with the ! symbol.

Execute a command, and display the output immediately:

> !pwd

This way you can also run commands like vim and top.

Capture STDOUT and STDERR

Capture the output of a command for later use:

> !!git diff

The standard output can be inserted into the chat with %STDOUT.

> Please write a commit message: %STDOUT

The results can be referenced through the environment variable RESP (experimental).

> !git commit -m "$RESP"

You can also insert the contents of the standard error output into the chat using %STDERR.

> !!wrong_command

> Explain the meaning of this error message: %STDERR

You can also use !{ cmd }. In this case, it will be replaced by the contents of standard input

> What time is it now? Hint: !{date}

Run Code Blocks in the Response

When ChatGPT returns code blocks, ChatGPT CLI saves these blocks temporarily and assigns them to environment variables named CODE1, CODE2, ... and so on. This allows you to execute the code blocks on your computer.

> Write code to display 1 to 10 in Python and Ruby.
> Python:

```
for i in range(1, 11):
    print(i)
```

Ruby:

```
(1..10).each { |i| puts i }
```

> ! python $CODE1
> ! ruby $CODE2

Save code block only.

> Please write comments/documentation for the code: %{src/client.cr}
> !cp $CODE1 src/client.cr
# Then check diff with your code editor...

Configuration

config.json is located in ~/.config/chatgpt-cli/ by default.

• substitute_patterns: %{} %%{} !{} %STDOUT %STDERR
• terminal_colors: chatgpt warning debug stdout stderr

Type %config to launch editor. Placeholders such as %{} and %%{} can be set with regular expressions

Please refer to the latest config.json file in the repository for the most recent information.

Prompts (system commands)

prompts.csv is located in ~/.config/chatgpt-cli/ by default.

Uninstallation

ChatGPT CLI uses the following 3 files and directories. This is all there is to it.

rm /usr/local/bin/chatgpt   # Remove the executable
rm -r ~/.config/chatgpt-cli # Remove the config directory
rm ~/.chatgpt_history       # Remove command history

Or

sudo make uninstall

Contributing

Bug reports and pull requests are welcome on GitHub at https://github.com/kojix2/chatgpt-cli.

Development

Building:

git clone https://github.com/kojix2/chatgpt-cli
git submodule update -i
cd chatgpt-cli
shards install
shards build

Testing:

crystal spec

You can quickly try development by using Gitpod.

License

This project is open source and available under the terms of the MIT License.