New-dev0 / termux-voice-gpt

Using GPT-4 to remotely execute code on my smartphone is a smart idea, surely

ai-control v2.1

Using GPT-4 to remotely execute code on my smartphone is a smart idea, surely. Supports voice control, and loading OpenAI Plugins. Comes with Termux shortcuts.

What's new?

The assistant is now able to effectively execute multiple actions in a row. It does this by calling another LLM, which comes up with a plan. It also has access to a bunch more Termux API tools. Voice chat is now a mode, thanks to Termux Speech to Text, which uses your phone's engine. The Assistant can now load OpenAI Plugins to use.

Just added: partial gpt-3.5-turbo support!

Video Demo (YouTube)

features

Not all of these have been tested, so please open an issue if you find something not working.

Tools using Termux API:

• BatteryStatusTool
• BrightnessTool
• PhotoTool
• ClipboardGetTool
• ClipboardSetTool
• FingerprintTool
• LocationTool
• MediaPauseTool
• RecordMicTool
• NotificationTool
• ListNotificationsTool
• RemoveNotificationTool
• URLOpenerTool
• TorchTool
• SpeakTool
• GetVolumeTool
• SetVolumeTool
• WiFiInfoTool
• WiFiScanTool
• VibratorTool
• MediaPlayTool
• SearchContactsTool
  • This lists all contacts, but filters by name to avoid a prompt too large for OpenAIs limits.
• ListSMSTool
• SendSMSTool
• GetCellInfoTool
• StartCallTool
• ListSensorsTool
• ReadSensorTool

Utility Tools

• Sleep
• PlanTool (this is the second LLM that lets the assistant perforn complex series of commands)
• GoogleSearchTool - no API key needed

OpenAI Plugins

Langchain built their own plugin loader, but I found it didn't work how I wanted it to, so I built on top of that. You can see my work in PluginLoader.py. Plugins will be loaded by URL, from the list defined in your config which I will go over in the setup/installation section.

Future ideas

Ngl I'm writing this for myself. Sorry if I get too verbose with it.

• executing shell commands
• scheduling events for the future
• add an option, for when running chat from a Termux environment (or when running both - as it must be a Termux environment), to use Termux speech to text and vice versa instead of typing to the bot

setup/installation

Make a copy of .env.example named .env and set the following keys:

• OPENAI_API_KEY - from OpenAI
• TERMUX_AGENT_URL - the URL of where the termux agent is running
  • If you're running the termux agent and the controller on the same device, you can skip this step as it defaults to http://localhost:8080
  • Otherwise, it's just http:// + your phone's IP address + :8080 for the port

For the agent device, make sure you have the Termux-API app installed along with the interface. To install the interface, run pkg install termux-* from within a termux session. Be sure to enable all permissions you want to use for Termux-API - for Android 13+ you might need to allow restricted settings.

config.json

This is the default configuration file which is loaded when starting --chat mode. You can specify which to load with the --config <filename> argument. It should contain the following keys:

• tools - this is where you can define additional tools you want your Assistant to have. It should contain another dict with the following keys:
  • langchain - a list of tools included with langchain. If any of those tools require an API key, set it in your .env file.
  • openai - a list of URLs which should link to an OpenAI Plugin (example). This feature is highly experimental! Also, this requires you include the requests_all tool in the langchain section.
• model_name - here you can change the name of the LLM used by the chatbot. By default, for best results, gpt-4 is selected. You can change it to gpt-3.5-turbo however it is not recommended.
  • Please note, the PluginLoader tool used for loading OpenAI Plugins uses gpt-4 and cannot be changed.
  • Also, PlanTool in UtilTools uses text-davinci-003.

usage

Verbose mode will be set on as default so you can see what actions are performed. Also, for the Termux shortcut assistant-voice.sh, --verbose-voice is applied by default - causing the tools used to be spoken aloud.

1. Run main.py for the CLI interface

termux shortcuts

Requires the Termux-Widgets app

• Copy the .shortcuts folder to your home directory (or move the scripts inside, if you already have a .shortcuts folder in your home dir)