New Open-Source Browser Engine Aims to Cut Token Costs for AI Web Agents

Developers building autonomous browser agents often find that sending screenshots to vision language models consumes large token budgets and slows each step. A new tool called agent-browser-mesh, available via pip, is designed to address this by converting page DOM elements into compact selector references instead of images. Its developers say it transmits only DOM nodes that changed between steps, reducing per-step token use by about 92 percent compared with vision-based approaches. The project also includes a security check intended to flag prompt injection payloads hidden in HTML before they reach the model. The creators report that these features lower latency and remove the need for GPU resources in headless deployments.