Case Study
Browser-native AI Agent that drives the web
An AI agent in your browser's side panel that reads and operates any web page, runs on the model you choose, and asks before it acts.
Project Details
#AI Agents
#Browser Extension
Auto Browser is a Chrome extension we built in-house. It puts an AI agent in the browser's side panel that reads the page you are on and operates it for you: running a search, filling a form, working through a checkout, pulling a figure out of a dashboard. You describe the task in plain words and watch it work one step at a time.
Most automation needs a site to expose an API or a maintained integration first. Auto Browser does not. It reads the page's accessibility tree the way a screen reader does, so it works on sites that were never built with an agent in mind. When a page publishes its own tools through the WebMCP standard the agent uses those instead, and it ships that runtime itself so pages adopting the standard work straight away.
The model was never the hard part. Everything around it was: reading a messy live page reliably, noticing when an action quietly failed, recovering from a stuck state, and never doing something the user did not ask for.
You bring your own model. The same agent runs on OpenRouter, on Google Gemini, on a local server such as Ollama, or fully on-device through Chrome's built-in Gemini Nano with no network calls at all. There is no account to create, no analytics and no telemetry, and API keys stay on the machine.
An agent that can click “Buy” needs guardrails that hold. Every action that changes something is gated by a permission the user grants per site, and the page origin is captured at approval and checked again when the action fires, so a redirect between yes and click cannot retarget it. Banking, government and healthcare sites are blocked by default.
The work we are most proud of is the testing. We wrote harnesses that score how often the agent picks the right next action and that drive the real extension through full tasks, under one rule: no scorecard, no claim of improvement. We also audited our own architecture against the best-documented browser agents published this year, wrote down where we fell short, and published the gaps instead of papering over them. Auto Browser is live and free at version 1.3.0, backed by more than 1,400 tests, and still moving.
Most automation needs a site to expose an API or a maintained integration first. Auto Browser does not. It reads the page's accessibility tree the way a screen reader does, so it works on sites that were never built with an agent in mind. When a page publishes its own tools through the WebMCP standard the agent uses those instead, and it ships that runtime itself so pages adopting the standard work straight away.
The model was never the hard part. Everything around it was: reading a messy live page reliably, noticing when an action quietly failed, recovering from a stuck state, and never doing something the user did not ask for.
You bring your own model. The same agent runs on OpenRouter, on Google Gemini, on a local server such as Ollama, or fully on-device through Chrome's built-in Gemini Nano with no network calls at all. There is no account to create, no analytics and no telemetry, and API keys stay on the machine.
An agent that can click “Buy” needs guardrails that hold. Every action that changes something is gated by a permission the user grants per site, and the page origin is captured at approval and checked again when the action fires, so a redirect between yes and click cannot retarget it. Banking, government and healthcare sites are blocked by default.
The work we are most proud of is the testing. We wrote harnesses that score how often the agent picks the right next action and that drive the real extension through full tasks, under one rule: no scorecard, no claim of improvement. We also audited our own architecture against the best-documented browser agents published this year, wrote down where we fell short, and published the gaps instead of papering over them. Auto Browser is live and free at version 1.3.0, backed by more than 1,400 tests, and still moving.