---
schema: 4
type: case
id: 11
title: "A voice agent that browses the web for people who cannot use a screen"
description: "Speak a task, and an agent drives a real browser and talks the results back. Built with safety limits and an audit log for caregivers."
created: 2026-10-20
updated: 2026-10-20
attributes:
  channels: [website]
  status: published
  motif: voice
  kind: own
  client: "h1902y"
  year: 2026
  state: "Experiment"
  stack: ["Python", "FastAPI", "Playwright", "browser-use", "WebSockets"]
  timeline: "July 2026"
---

## Problem

Blind, low-vision and cognitive-disability users need to get things done on the web without a screen or an install.

## What we built

- A voice-first web channel with a browser speech client and a session state machine, including stop words and spoken confirmation
- A safety layer: blocked categories, actions that need confirmation, and rules per domain
- Usage caps per user and overall, and an append-only audit log for caregivers
- An agent runner that drives a real browser, plus a phone channel skeleton

## Outcome

- Working on localhost; not deployed
- Open risks are written down: browser isolation, privacy and prompt injection