What it can do
Switch on Computer in the message box, say what you need, and Laghu works through it the way a careful person would: look at the screen, take one step, look again to see that it worked.
- Websites. Open pages, search, fill in forms with text, lists and tick boxes, go through pages of results, switch between tabs and answer a page’s pop-up questions. It uses a browser window of its own, beside the one you normally use.
- Windows apps. It works with an app’s buttons, menus, fields and lists through Windows’ accessibility features, the ones screen readers use. So it presses the actual “Save” button, not a spot where a button might be.
- Finding and reporting. It keeps notes as it goes, so a fact found on the third page is still known on the tenth, and it ends with the answer or a summary of what it did.
Things to ask it
- “Open the delivery page on this shop’s site and tell me how much delivery costs.”
- “Fill in the contact form at this address with my name and this message, and stop before sending.”
- “Search the library catalogue for books by Ruskin Bond and list the ones that are available.”
- “Go through the first three pages of results and note every listing under ₹20,000.”
- “In Notepad, write a short leave letter from these notes and save it on the Desktop.”
Clear tasks with a clear end go best. Say where to start, a web address if you have one, and when to stop.
How you stay in charge
- You agree to each task. Nothing happens until you press Start on a card that repeats what you asked for.
- It never types your secrets. When a field asks for a password, a card number or a one-time code, Laghu stops and hands over. You type it in the browser window yourself, press Done, continue, and it carries on. What the model suggested typing there is not shown or kept.
- It asks before anything that matters. Paying, buying, sending, submitting a form, deleting, booking, subscribing or ticking “I agree” waits for your yes, every time, whatever you allowed before. If you say no, it is told not to look for another way to do it.
- You can see it working. While it works, a soft blue light runs round the edges of your screen with a small label at the top saying what it is doing. The light turns gold when it is your turn. It is left out of the pictures the model sees.
- You can stop it from anywhere. Press Ctrl+Alt+End, whatever window is in front, or press Stop in Laghu.
- It does not fight you for the mouse. When you use the mouse or keyboard, it waits until you have finished before it clicks or types in an app.
- Some things are off limits. Laghu’s own window, command windows, and Windows’ security and system tools such as Task Manager and the sign-in prompt. Downloads are switched off.
- It keeps a record. Every step and every answer you gave goes into the activity log in the Agent screen.
It can still misread a page, so watch the first few tasks, and keep the yes for steps you have checked.
Start in a minute
- Make sure Google Chrome or Microsoft Edge is installed. Laghu opens its own window of it for websites.
- In the model picker, choose a model that sees. Qwen2.5-VL 3B is in Laghu’s library, about 2.8 GB, and is the best choice on an everyday laptop. With Computer on, the picker lists models that see first and offers it if you have none.
- Switch on Computer in the message box.
- Describe the task: where to start, what to do, and when to stop.
- Press Start on the card, and answer any card that asks for you.
As it works you see each step, with the picture of the screen it decided on, in the conversation.
Which model to use
A model that sees is best, because it gets a picture of the screen as well as a list of what is on it, and so understands icons, images and layouts. Qwen2.5-VL 3B is a good choice on an everyday laptop, and Qwen2.5-VL 7B reads dense pages better at about half the speed; both are in Laghu’s library, and a copy you already have in Ollama works too. The 3B’s licence allows research and personal use, the 7B’s commercial use as well. SmolVLM2 500M and 2.2B are smaller and quicker, and better kept to short tasks. A model that reads text only, such as Qwen2.5 7B, can still work from the list of what is on the page.
Each step is a full turn of the model, so on a laptop a task of ten steps can take several minutes. That is by design: the task is the kind you would otherwise spend those minutes clicking through yourself, and every step is checked. What speed to expect.
Tips for good results
- Give it a starting address. “Go to this page and…” saves it a search.
- Say when to stop. “Stop before paying” or “don’t send it” makes the end clear.
- Sign in when it asks. Its browser window is its own, so it starts signed out. When a site needs your account, it hands you that step.
- One errand at a time. Several short tasks go better than one long one.
Questions
Can Laghu fill in forms on websites for me?
Yes. It types into fields, chooses from lists and ticks boxes in its own browser window. Submitting a form waits for your yes, and it never types passwords, card numbers or one-time codes; it hands those to you.
Does anything from my screen leave my computer?
The pictures it takes of your screen go only to the model on your PC. The websites it opens see the visits, as they would if you opened them.
Which apps can it use?
Most Windows apps, through the accessibility features they offer to screen readers. Laghu’s own window, command windows and Windows’ security and system tools are off limits.
How do I stop it?
Press Ctrl+Alt+End from any window, or Stop in Laghu. It also stops whenever you say no to starting, or press Stop the task on a card.
Is it fast?
Not especially, on purpose. Each step looks at the screen and checks the last one, and on a laptop that takes a little while. It is meant for errands you can hand over and come back to.