open-ingress/openingress

OpenIngress checks whether an AI agent, the kind that browses websites on someone's behalf by clicking buttons and filling in forms, can actually use your site. A page can look completely normal to a human in a browser while still confusing an agent, for example if buttons have no labels, if content only appears after JavaScript runs, or if important links never show up in the underlying page structure that agents read. The project points out that being easy to find through search engines is a different problem from being usable once an agent lands on the page. To test this, OpenIngress crawls a site the way a real browser would, using Playwright, following links up to a set depth and page limit. While crawling, it captures the page's HTML, screenshots, and an accessibility snapshot, which is a structured view of what an agent would actually be able to perceive and interact with. From that data it produces operability scores, checks whether important content loads properly, and builds a map of how pages connect to each other. Beyond this static analysis, the tool can also run exploration tasks using a large language model of your choice, connected through an OpenAI-compatible API. These tasks attempt to actually navigate the site as an agent would, and the results are recorded as traces showing exactly where an agent got stuck or blocked. All of this comes together in a final report showing which pages were reached, what broke along the way, and where navigation failed. The project is self-hosted, with a Flask backend API and a Vue based web dashboard for viewing results. Getting started involves copying example environment files, providing an API key for the language model you want to use, and running a setup command that installs Python and Node dependencies along with the Playwright browser. It is released under the MIT license.