Automate Browser and Desktop Chores for €0 With AutoCua, an Open-Source Agent
In short
- AutoCua is a free, MIT-licensed agent that drives your browser, desktop apps and terminal with an AI model of your choice. It showed up on Hacker News on November 4.
- You pay only for the model API. It works with nine providers, including OpenAI, Anthropic, Google and OpenRouter.
- The money angle: a 30-minute daily click job costs about €88 a month in time. The same run on an API is a few euros.
You have a task you repeat every day. Open a site, copy numbers, paste them somewhere else. A person costs money. A script breaks when the page changes. A computer-use agent looks at the screen like you do and clicks. AutoCua gives you one for free.
What it is
AutoCua is an open-source computer-use agent. The repo describes it as an automatic agent for end-to-end automation workflows on macOS, Windows, Linux, iOS and the browser. A language model looks at the screen, decides what to do, and the tool clicks, types and runs commands for it.
It has five modes: computer use for desktop apps, browser use, shell use, web use (it drives your own Chrome through an extension) and mobile use for iPhone and iPad simulators. Mobile is macOS only, and Android is not available yet.
Who is behind it? The repo sits under the ComputerUseAgent account and I could not find a company page. That is normal for a new open-source project, but remember it when you decide what to give it access to.
What it costs
| Part | Cost |
| AutoCua itself | Free, MIT license |
| Model API | Pay per token to your provider (OpenAI, Anthropic, Google, OpenRouter, Groq, Cerebras, Together AI, Perplexity, AWS Bedrock) |
| Windows UAC driver (Interception) | Separate step. Dual-licensed: commercial use needs a paid license from its author |
| iOS mode | Needs full Xcode on a Mac |
The repo lists 48 model names in a file called model_list.txt. Names are case-sensitive and not validated, so a wrong name gives you a 404.
Do the math
Do the math: take one job: 30 minutes a day of copying data between two web pages. At €8 an hour that is €4 a day, or €88 a month over 22 working days. Now give it to the agent. Assume one run costs €0.20 in model tokens (my assumption, it depends on the model and the page). That is €4.40 a month. You save roughly €83 a month on one job, about €1,000 a year. Ten jobs like this and the number changes how you plan hiring. Test your own run cost before you believe my figure.
Set it up in 20 minutes
- Check you have Python 3.10 or newer. Run
python --version. - Install the package:
pip install AutoCua. On Linux, first install the distro packages the README lists (GNOME desktop, AT-SPI, GTK bindings). It was tested on Ubuntu with Wayland, X11 is not tested yet. - Get an API key. OpenRouter is the easiest start, one key and many models. You can also use OpenAI, Anthropic or Google directly.
- If you work from a checkout, run the setup script for your system (setup_mac.sh, setup_windows.bat or setup_linux.sh), copy .env.example to .env and paste your key there.
- Pick a model name from model_list.txt. Copy it exactly.
- Write the task in plain words and run it with the launcher below.
- Start with a read-only task such as summarising a page. Only then try tasks that type or click.
from AutoCua.agent_launcher import run_agent
run_agent(mode="web use", provider="openrouter", model="gemini-3.8-flash",
task="Open wikipedia.org and summarise today's featured article.")Then run python main.py. For several tasks at once, the README supports parallel runs through extra_tasks=[...].
One thing to know for browser work: Google Chrome is needed, and AutoCua installs it on first run if it is missing.
Real use cases
- Price checks. Open three supplier portals, read the price of the same item, write it into a list. A parts shop like GiantParts has this kind of job all day. I plan to test it there on a read-only task first.
- Report copy-paste. Pull numbers from a dashboard that has no export button and put them in a sheet.
- Form filling. Fill the same supplier or marketplace form from a table of data.
- Testing your own site. Tell it to walk through checkout and report what broke.
Limits and weak spots
Read this part before you install anything. The README is honest about it.
- The shell tool is not a sandbox, and there is no approval prompt before commands run. A wrong instruction can delete or send things. Use a virtual machine or a spare computer.
- There is no automated test suite in the repo.
- Read-only behaviour for shell commands is enforced by prompt, not by code. That means a model can ignore it.
- Some web search fallbacks (Together AI, Cerebras, AWS Bedrock) run through the browser agent and can take minutes.
- The Discord and WhatsApp integrations are placeholders and do not work.
- Websites that look for the browser overlay can detect it.
- Free Apple ID provisioning profiles for real iOS devices expire after 7 days.
My honest read: it is a young project. The star count was low when I looked at the repo page, and other sites quote different numbers, so I would not trust popularity as proof. Judge it by your own test.
Alternatives
- Official computer-use tools from the big labs. More polish and support, but you pay their price and live in their rules.
- Browser-automation libraries. Good if the task lives only in a browser and you can write a little code.
- Classic automation tools like Make or n8n. Best when the site has an API. Always cheaper and more stable than an agent that clicks, so try them first.
Ivan’s Take
Worth testing, not worth trusting yet. A free agent that works on desktop, browser and phone simulators is rare, and the math on repeat jobs is strong. But no sandbox and no approval step means I would run it only inside a virtual machine, with a throwaway API key, on read-only jobs first. If it passes a week of that, I will move one real supplier-price job to it.
Source: AutoCua on GitHub
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
