AIVA Agentic Browser · Early access

The browser that
works for you.

AIVA Browser has an agent built in. Today it can navigate to a page on request and read or summarize what's on it. Clicking, filling forms, and chaining multi-step tasks are actively in development.

Windows beta available now Public source on GitHub No account needed No payment info requested
Solar power — Wikipedia
en.wikipedia.org/wiki/Solar_power
ENERGY
Solar power

Why AIVA

Browsers got faster.
They never got smarter.

Every tab you open is still a task you finish yourself — click by click, form by form. AIVA is being built to turn intent into action, without handing over control of what happens on your machine.

Hours back, not tabs

Delegate the parts of browsing that don't need you — starting today with reading and summarizing pages, with clicking and form-filling actively in development.

Built with limits in mind

The agent is instructed to never enter payments or credentials without asking. A hard, enforced confirmation step before consequential actions is on the roadmap.

Still just a browser

Fast, familiar, Chromium-based. The agent is a layer you can lean on or ignore — it never gets in the way of ordinary browsing.

The core differentiator

It reads you, not just the page.

Other agentic browsers read the page they're on and act on it. AIVA also reads the person using it. The face scanner — built for fatigue detection — keeps a rolling 15-minute picture of your strain, screen time, and time since your last break, and feeds it straight into the same agent that's driving your browser.

Available now

The agent interrupts itself

Past step 5 of a long task, if you've been visibly worn out for a sustained stretch, the agent stops and asks — keep going, or pause for a break — instead of quietly burning through the rest of its step budget.

Available now

It answers differently, not just more

At high strain, responses get shorter and skip optional side-quests. It never diagnoses you, and mentions your state at most once per conversation — never repeatedly.

Available now

Camera-driven nudges, not fake timers

Posture is genuinely measured: AIVA learns your neutral sitting position over the first 20 seconds, then flags real drift — slouching, leaning in, a tilted head — after 45 seconds sustained. Eye-care reminders follow the real 20-20-20 rule based on actual detected screen time. Only water stays a plain timer — no webcam detects thirst.

Available now

It offers, never forces

Breathing, neck-exercise, and meditation break suggestions arrive as a card with a button. Your work is never touched, and nothing navigates without your click.

Mid-run wellness checkpoint
CHECKPOINT — STEP 6

"You've looked drowsy for a few minutes straight. Keep going, or pause and take a break?"

Keep going Pause & take a break

Other agentic browsers finish your task. AIVA finishes it while paying attention to whether you should still be doing it.

How it works

Where the agent stands today.

01

Tell it the task

Type a goal in plain language, like "open this article and summarize it."

02

It navigates & reads

AIVA can go to a page on request and read its visible content — available today.

03

It's learning to act

Clicking, filling forms, and chaining multi-step actions across a task are actively in development.

04

You'll stay in control

An enforced confirmation step before consequential actions, plus an audit log of everything the agent does, is planned.

Features

What's real today —
and what's next.

Every card below is labeled honestly. Nothing is marked "available now" unless it's actually shipped.

Available now

Read & summarize the page

Ask AIVA about the page you're on, or tell it to open a URL — it reads the visible content and can summarize or answer questions about it, narrating each navigation step.

Available now

Stop, anytime

A visible Stop control and a 14-step hard cap end an in-progress run instantly — no silent, open-ended automation.

In development

Clicking & form-filling

Filling inputs the way a real user types, so modern framework-driven forms register the change correctly. Not yet reliable enough to call finished.

In development

Multi-step task chains

Chaining navigation, reading, clicking, and filling into one instruction the agent plans and executes across several steps.

navigate read_page click_element fill_input
Available now

Wellness, built in

Opt-in fatigue detection, posture tracking, and guided breaks sit alongside the agent — the camera stays off until you turn it on.

In development

Multi-tab awareness

Opening, switching, and closing tabs as part of a task, so the agent can work across more than one page.

Security & trust

An agent with hands
should have limits.

Here's exactly what's enforced today, what's only an instruction to the model, and what's still on the roadmap.

  • ✓

    Page content is reference, not instructions Today

    The current page's visible text and URL are sent to the model only as context — never executed or treated as a command.

  • ✓

    Bounded, cancellable runs Today

    Every agent run has a hard 14-step cap and a visible Stop control — no silent, open-ended automation.

  • !

    Told not to touch payments or credentials Instruction, not enforced

    The system prompt tells the agent never to enter payment details or credentials. That's model guidance today — not yet a hard, code-enforced block.

  • ○

    Enforced approval & audit log Planned

    A UI confirmation step before consequential actions, and an audit log of everything the agent does, are planned but not built yet.

  • ✓

    Public source on GitHub Today

    The code is public and viewable today. A formal open-source license hasn't been published yet, so treat it as source-available for now.

backend/server.py — system prompt excerpt
// actual instruction given to the model
"Do not enter passwords, payment
 details, or other sensitive personal
 data without explicit confirmation."

[status] model instruction — active today
[status] code-enforced block — planned

Download

Get AIVA Browser for Windows.

v0.1.4 is our latest stable release — still unsigned, so Windows SmartScreen may warn on first run; click "More info" → "Run anyway" if you trust the source. macOS and Linux builds are in progress.

v0.1.4Windows

v0.1.4 · ≈198 MB · Windows 10/11 (64-bit) · or run it from source

Demo video — recording in progress.

We'll publish a captioned walkthrough with a full text transcript here. Until then, here's what AIVA can do today: navigate to a page you name, and read or summarize what's on it.

Status

Available now, in development, and planned.

Built in the open — this list changes as work lands, not as a marketing promise.

Available now
  • Navigate to a URL on request
  • Read & summarize the current page
  • Stop any in-progress agent run (14-step cap)
  • Wellness tools — fatigue, posture, guided breaks (opt-in)
In development
  • Clicking & filling forms on a page
  • Multi-step task chains across tools
  • Multi-tab operation (open, switch, close)
Planned
  • Enforced permission prompts before consequential actions
  • Audit log of everything the agent does
  • Reusable, saved workflows

FAQ

Common questions

Today: navigate to a page you specify, and read or summarize what's on it. In development: clicking, filling forms, and chaining multi-step tasks. Planned: enforced permission prompts and an audit log before the agent takes consequential actions.
Yes. AIVA is a full Chromium-based browser with tabs and browsing history — the agent is an added layer, not a separate app you switch to.
Not today — payment and account flows aren't wired up at all, and the model is instructed never to enter credentials. A hard, code-enforced confirmation step is planned, not yet built. You can stop any run instantly with the Stop control.
Only what's needed for the task at hand: the current page's visible text and URL when you're chatting with the agent, plus the tool actions it takes. See our Privacy Policy for full details.
The code is public on GitHub today, but we haven't published a formal open-source license yet — treat it as source-available for now. Publishing a real license is next on our list.
Windows: today, as an early beta — see the Download section. It's unsigned, so SmartScreen may warn on first run. macOS and Linux don't have a packaged build yet; watch the GitHub repo for release updates.

Curious where this goes next?