OpenTARS

agent
Security Audit
Fail
Health Warn
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 6 GitHub stars
Code Fail
  • rm -rf — Recursive force deletion command in app/empacotamento/appimage/build.sh
  • rm -rf — Recursive force deletion command in app/empacotamento/appimage/dentro_do_chroot.sh
  • rm -rf — Recursive force deletion command in app/empacotamento/build.sh
  • rm -rf — Recursive force deletion command in app/empacotamento/linux/instalar.sh
  • rm -rf — Recursive force deletion command in app/empacotamento/macos/instalar.sh
  • rm -rf — Recursive force deletion command in app/empacotamento/setup.sh
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Local AI assistant for Linux, Windows and macOS. It operates your computer from plain language, with its own 617 MB model. Free, open source, local by default.

README.md
openTARS logo

openTARS

Your AI, on your computer.

Ask in your own words. openTARS opens programs, clicks buttons, types, finds files, reads the screen and runs commands.

Local by default, free and open source, with an AI trained just for it.

Version 5.3 Tesseract
Official AI: Gargantua 2.5
MIT License
Linux, Windows, macOS
Languages
Instagram @open.tars

Install · What it does · Gargantua · How it works · Safety · Terminal · New in 5.3 · Website


The openTARS quick bar, opened from anywhere: do a task, ask about the screen, find a file, search the web or create a reminder.

Install

One command. It sets up everything, downloads the models and tests them.

Windows 10 or 11, in PowerShell:

irm https://raw.githubusercontent.com/MastLLM/OpenTARS/main/app/empacotamento/windows/instalar.ps1 | iex

Or download opentars-setup.exe.

macOS 12 or newer, in Terminal:

curl -fsSL https://raw.githubusercontent.com/MastLLM/OpenTARS/main/app/empacotamento/macos/instalar.sh | bash

Or download the app for Apple Silicon or Intel.

Ubuntu, Debian, Mint, Zorin, Pop!_OS:

wget -O /tmp/opentars.deb https://github.com/MastLLM/OpenTARS/releases/latest/download/opentars_all.deb && sudo apt install -y --reinstall /tmp/opentars.deb

Fedora, Arch, openSUSE:

curl -fsSL https://raw.githubusercontent.com/MastLLM/OpenTARS/main/app/empacotamento/linux/instalar.sh | bash

Raspberry Pi, Orange Pi and other small boards: see openTARS Pi, the edition made for them.

curl -fsSL https://github.com/MastLLM/OpenTARS-Pi/releases/latest/download/install.sh | bash

Then open openTARS from your menu. Windows and macOS need no administrator password: everything goes into your user folder. A short first-run window picks the right model for your computer. Running the same command again updates, and your chats and settings stay.


The openTARS window: history sidebar, the request, the steps card, tool calls with timings and the answer with a code block

What it does

Work on your computer in plain words. "Close Spotify and open Discord." "Organize my Downloads folder by type." "Put Firefox on the right half." It opens and closes programs, moves windows, types, clicks buttons by the name they show, and works with your files.

The quick bar with nothing typed, offering: do a task on the computer, ask about the screen, find a file, search the web and create a reminder The text actions panel rewriting a selected sentence, with Replace in the text, Copy and Discard Editing a selected photo: square and brighter, with a before and after preview and the button to save a copy
The launcher, from anywhere Text actions in any app A photo edited by sentence
🚀 Launcher Ctrl+Alt+Space from anywhere: do a task, ask about the screen, find a file, search the web or create a reminder, without switching windows
✍️ Text actions in any app Select text and press Ctrl+Alt+R (Control+Option+R on a Mac): fix it, make it more formal or simpler, summarize, translate between English and Portuguese, or ask about it. Replace puts the result right back
👀 It sees your screen "What is on my screen?" Gargantua reads it too, and finds icons that have no text
🔎 Find anything Searches names, folders and the text inside your documents, PDF included
📂 Actions on files Summarize, rename by content with a preview you can edit, move, or send to the trash after you confirm
🖼️ Photos by sentence "Crop it square and make it brighter." Remove the background, upscale 4x, or click an object to remove it. The result is always a new copy
🎨 Make images "Draw a lighthouse at dusk," with the optional Bonsai Image 4B: about 5 s per 512-pixel image on a 6 GB card, after the first one
🎙️ Voice Hold to talk, continuous, or the wake word "TARS". Gargantua 2.5 understands what you say directly, in English and Portuguese, with Whisper as the backup
🛰️ Probes Small watchers in the background that tell you when something matters: a folder, a web page, news, a GitHub project, this computer
⏰ Routines and reminders "Every day at 8, open the newspaper." Created with a sentence
🧠 Memory "Remember my browser is Brave." It applies from then on, and you see and erase everything in Settings
🌐 Web search Reads the results and answers, through your own SearXNG, without opening anything on your screen
🖥️ Missions in a VM On Linux, Gargantua Pro works inside a disposable virtual machine. Watch it live and choose which files come back
📅 Calendar, notes, e-mail, music By sentence, all off until you turn them on. E-mail never sends without your yes
🛍️ Model store Hand-picked models on Hugging Face, or search thousands of GGUF models; every file checked by SHA-256

KIPP's moods: idle, thinking, working, needs you, oops, done and napping
KIPP walks along the chat and shows what the AI is doing

Gargantua, an AI made for this

Other assistants wrap a general model in a long prompt. Gargantua was trained on thousands of requests actually solved inside openTARS and checked one by one: the calculator showed the right sum, the folder existed, nothing was deleted. It already knows the tools, so it acts on the first try, with a short prompt, and it fits on modest hardware.

All of them run on Gemma 4, Google's open model family that takes in images and audio, except Code (Qwen2.5-Coder).

Size Runs on In our tests
Gargantua 2.5 5.3 GB a graphics card, or 16 GB of RAM the right tool 21 of 21 times, 75 of 76 text actions, 47 of 48 everyday tasks, about 1 s per tool on a graphics card
Gargantua Lite Gen 3 3.4 GB any computer, on the processor alone 75 of 76 text actions, 48 of 48 everyday tasks
Gargantua Pro Gen 1 7.4 GB optional, for missions in a virtual machine our largest model, on Gemma 4 12B
Gargantua Code Gen 2 4.7 GB optional, for programming 51 of 60 coding tasks
Gargantua Pi Gen 1 300 MB small boards, in openTARS Pi tested on an Orange Pi Zero 2W with 1 GB of RAM
Murph 4 MB always on, no model to load reads each request in about 0.2 ms and picks who answers

The tests are ours and run with the real app, on November 13, 2026; opentars --prova runs the everyday tasks on your computer. The models run on the llama.cpp engine that openTARS installs, and are also on Hugging Face and ollama.com.

How it works

flowchart LR
    P([Your request]) --> L{{"Murph and the layers<br/>keywords · context · format · apps"}}
    L --> M["Model queue<br/>Gargantua first for actions<br/>size · memory · success on your PC"]
    M --> IA["The AI<br/>plans when there are several steps"]
    IA --> F["Tools<br/>accessibility · vision · keyboard · terminal · web"]
    F -- "nothing worked" --> M
    F --> S([Answer])
  • It picks the AI by itself. Each request is read in milliseconds and goes to the best model you have, and the queue learns from what works on your PC.
  • Clicks by name, not by position. Buttons are pressed through the system's accessibility, with no screenshot and no coordinates. Looking at the screen is the plan B, and every click is checked afterwards.
  • "Done!" has to be true. If the AI says it did something without doing it, it is told to really do it. A task with several steps becomes a plan, and it cannot finish until every step is done.
  • It does not walk in circles. Repeating what already failed does not run again, and after repeated failures the next model takes over the same conversation.
  • One conversation. Switching AI midway does not lose what you asked.

Your computer, your rules

  • Local by default. With a local model, your conversation, your voice and your screen stay on your machine. Cloud models and web search are optional and go only where you point them.
  • It asks first. Deleting, sending and installing wait for your Allow. It never closes a program with unsaved work.
  • Routines and Probes stay on a leash. A routine only does what you allowed when you created it, and a Probe only reads unless you chose an action.
  • No account, no subscription. Keys you add stay on this computer and are never shown or logged. It also checks GitHub for updates, and you can turn that off.

More in the privacy page.

Terminal

opentars opens the chat right in the terminal, on a desktop or over SSH: you type in a framed prompt with a status bar under it (model, engine, folder and git branch), follow each step as it runs, and read formatted answers. Risky commands ask first.

The openTARS terminal: the logo, the welcome panel, a plan, tool steps with their results and the spinner
In the chat What it does
/help lists the commands
/models, /model <name>, /model auto lists, pins or unpins a model
/resume, /compact, /context continues a saved chat, summarizes it, shows how much is in use
/mode ask | plan | auto asks first, only describes, or applies edits in this folder. Deleting, sending and installing always ask
/undo reverts the file edits of the last request
!<command> runs a command and passes its output to the next request

In scripts: opentars -p "request" prints only the answer, --output-format json gives it as data, and git diff | opentars -p "review this" reads what you pipe in.

Command What it does
opentars-gui opens the window
opentars --diagnostico checks your setup and saves a report you can send us
opentars --prova runs everyday tasks with your model in a temporary folder
opentars --explicar "request" shows how the task and the model were chosen
opentars --searxng sets up web search on this computer, with Docker
opentars --desinstalar removes openTARS on any system, and asks before deleting your data or models
opentars --help everything else

Where it runs

Linux Ubuntu 22.04+, Debian 12+, Mint, Zorin, Pop!_OS, Fedora, Arch, openSUSE. GNOME, KDE, Cinnamon, XFCE, MATE and others. X11 has everything; on Wayland, chat, apps, commands, screenshots and clicking by name work
Windows 10 and 11. Voice is the newest part here
macOS 12 or newer, tested so far on GitHub's macOS machines. The first run asks for Accessibility and Screen Recording
Small boards 64-bit boards with openTARS Pi
No screen SSH, servers and WSL with opentars --chat
Graphics card NVIDIA, AMD or none

Languages

The window, the terminal, the installer and the menu speak English and Portuguese. Switch in the window or with opentars --idioma pt_BR. To add your language, copy app/idiomas/en.json, translate the values and send a pull request: it shows up in the menu by itself.

New in 5.3 Tesseract

  • Rebuilt on llama.cpp. One fast engine for every model, installed by the app the first time it opens. No separate AI program, no password.
  • Gargantua 2.5 on Gemma 4. The new main AI sees your screen and hears your voice.
  • Voice, understood directly. What you say goes straight to Gargantua 2.5.
  • Missions in a virtual machine. On Linux with KVM, Gargantua Pro works in a disposable computer, from a new window in Probes.
  • Native apps for Windows and macOS, with Python, Qt and the engine inside.
  • A model store on Hugging Face, with search and SHA-256 checks.
  • A terminal that looks the part.
  • One-click updates on Windows, macOS and every Linux.
  • Clicks in Chromium and Electron apps, reopened with accessibility on when you allow it.
  • Calendar, notes, e-mail, music and system settings by sentence, on Linux, Windows and macOS.
  • Verify your download with the checksum and the GitHub attestation of each release.

Help

  • Something is off? Run opentars --diagnostico and send us the report it saves, by opening an issue or at [email protected].
  • The AI does not answer: opentars --diagnostico checks the llama.cpp engine and the models, and says what is missing.
  • A shortcut does nothing: opentars --atalho shows what is registered, and opentars --atalho ctrl+alt+o picks another.
  • Clicking by name misses an app's buttons: open that app through openTARS, or turn on "Let openTARS reopen an app so it can click its buttons" in Settings → Assistant, for Chromium and Electron apps. On Wayland, an Xorg session gives clicks by position.

For developers

The app is in app/, and opentars_all.deb is the same release packaged for Ubuntu and Debian. The tests run on every change on Linux, Windows and macOS through GitHub Actions.

cd app
python3 -m tars_qt              # the window
python3 tars.py                 # the terminal
cd tests && python3 run_all.py  # the tests
bash empacotamento/build.sh     # builds the .deb

tars.py loads the core from nucleo/, tars_qt/ is the window, and tars_plataforma/ is the one door to what changes between systems. Gargantua's prompt and tools are exactly what it saw in training, and tests/test_gargantua_v40.py checks that byte by byte.

Follow along

News and what is coming: @open.tars on Instagram and the changelog.

License

MIT: you may use, copy, modify and distribute openTARS, including in other projects, as long as you keep the copyright notice and license with it. The software is provided "as is", without warranty of any kind.

Reviews (0)

No results found