Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

https://github.com/chrisdothtml/gpt-cmd

Sit back and let ChatGPT run your commands for you.
https://github.com/chrisdothtml/gpt-cmd

chatgpt cli command-line-tool llm task-scripts

Last synced: 16 days ago
JSON representation

Sit back and let ChatGPT run your commands for you.

Lists

README

        

# gpt-cmd

> Sit back and let ChatGPT run your commands for you.

> [!WARNING]
> It's always risky to execute commands directly from a third party. Use responsibly and at your own risk.

```sh
gpt_cmd "Install python3 and pip3 and symlink them to python and pip"

gpt_cmd "Install SQLite and write a script to verify it's working correctly"

gpt_cmd "Install black and generate a simple starter config file for my python project"
```

`gpt_cmd` lets ChatGPT iteratively run commands for you to achieve your end goal.

The way it works is that you provide some end goal (e.g. 'Install python3 and pip3'), the tool will bake-in information about your OS and architecture, and ChatGPT will respond with a list of commands it wants to run. The tool will execute these commands and respond to ChatGPT with the stdout and exit code of each. Once ChatGPT thinks it's done, it'll respond accordingly and the loop will end (and it'll usually provide a bit of context about what it did).

With this approach, ChatGPT is able to probe your system and try running commands, responding to potential failures as they happen with alternative commands. It's especially handy for tasks that aren't all that complex, but you can't be bothered to google exactly what you need to run.

## Install

> [!WARNING]
> In light of my other warning above, this install script is pulled directly from my GitHub repo, and is a potential vulnerability if the repo (or GitHub) becomes compromised. Always inspect scripts for shady behavior before running them on your device (even mine: [install.sh](https://raw.githubusercontent.com/chrisdothtml/gpt-cmd/main/install.sh)).

### Linux/MacOS

**NOTE**: the only system requirements are `bash` and either `curl` or `wget`.

```sh
curl -s https://raw.githubusercontent.com/chrisdothtml/gpt-cmd/main/install.sh | bash

# or if you prefer wget
wget -qO- https://raw.githubusercontent.com/chrisdothtml/gpt-cmd/main/install.sh | bash
```

The install script will make its best attempt to expose the binary to your `$PATH`, but if that doesn't work, you'll have to manually add it your path (binary install location is `$HOME/.gpt_cmd/bin`).

### Windows

There's not currently an automated installer for Windows, but you can download the `.exe` file from the [releases page](https://github.com/chrisdothtml/gpt-cmd/releases).

## Use

**NOTE**: before running, you need to create an `~/OPENAI_TOKEN` file and put your token in it.

```sh
gpt_cmd

# see `Env var overrides` section below for full list
GPT_CMD_MODEL="gpt-4-turbo" gpt_cmd
GPT_CMD_TOKEN_FILE_PATH="/my/token/file" gpt_cmd

# print path to the dir containing message logs for your previous runs
gpt_cmd --get-convos-dir
```

The `goal` can be literally anything you can achieve via a terminal (which is a lot). Even if it takes dozens of commands to get there, it'll eventually get there (maybe). You can be as descriptive or vague as you want, and list as many different tasks as you want (e.g. 'Install [some tool] and then write a starter script with some examples of how to use it').

## Env var overrides

Enironment vars that you can provide to change the behavior of the tool.

### `GPT_CMD_MODEL`

Override the gpt model used by the tool.

**Default**: `gpt-4o`

### `GPT_CMD_TOKEN`

Provide your OpenAI token via this env var instead of storing it in a file.

### `GPT_CMD_TOKEN_FILE_PATH`

Override the file path the tool gets your OpenAI token from. Alternatively, you can provide the token directly via `GPT_CMD_TOKEN`.

**Default**: `~/OPENAI_TOKEN`

### `GPT_CMD_DANGEROUSLY_SKIP_PROMPTS`

By default, the tool will prompt you before running each command, as giving an AI unfettered access to run commands on your system is usually a bad idea. However, if you're on a throwaway machine/container and don't care what happens to it, you can disable the prompts via `GPT_CMD_DANGEROUSLY_SKIP_PROMPTS=true`.

## License

[MIT](license)