Your AI agent can plan anything. It still can't click "Post".

Developer Wouter van der Stoep has introduced Ghost Browser, an open-source tool designed to give AI agents the ability to interact with the web like humans. While many AI agents excel at planning and reasoning, they often struggle with executing actions on websites due to brittle CSS selectors and authentication hurdles. Ghost Browser solves this by using a "Set-of-Mark" approach, where the agent identifies interactive elements via numbered boxes on a screenshot rather than parsing the DOM. This allows the agent to operate within a persistent, logged-in browser session, effectively bypassing common automation failures. The tool includes safety features like an "act-gate" that requires human approval for sensitive actions, and it supports residential IP routing to maintain session integrity. The project is currently available on GitHub, with the developer seeking community contributions to improve its user interface.