Gaze
A command-line tool that drives the browser you are already signed in to, and asks before anything on the page changes.
Problem
Browser automation almost always starts from a fresh, anonymous profile, so the first thing it needs from you is your passwords. I wanted the opposite arrangement, where the tool inherits the sessions I already have and the open question becomes what it is allowed to do rather than how it gets in.
Outcome
Open source and working against both engines. Reads are free, writes stop for approval, and it scores 100 out of 100 on a twelve-level obstacle course I built for it, based on published research on agent failure modes, including a prompt-injection level built to hijack the agent reading it.
Details
- Stack
- Node.js, Chrome DevTools Protocol, WebDriver BiDi, MCP
- Parts
- CLI, MCP server
- Links
- Source (open source)
Notes
Gaze keeps a clone of an everyday browser profile, so the logins come with it and nothing has to be handed over in plain text. Reading a page costs nothing and needs no ceremony. Anything that changes state stops and asks first, and that gate is enforced by the tool rather than left to the caller’s good manners, which matters most when the caller is an AI agent over MCP instead of a person at a shell.
The same rule covers the case people usually forget: a page can try to talk the agent reading it into doing something. On the benchmark’s prompt-injection level the answer is the same as everywhere else. The instruction is on the page, so it is data, not a command.