MolmoWeb

MolmoWeb is an open visual web agent that performs browser tasks using only screenshot inputs. It includes a full codebase with training scripts, an evaluation harness, an annotation tool for recording human demonstrations, a synthetic data generation pipeline, and client‑side demo UI. The accompanying MolmoWebMix dataset provides a large public collection of web‑navigation examples for training and benchmarking.

The system is aimed at researchers and developers who want to build, fine‑tune, or evaluate web‑automation agents without relying on proprietary models or hidden data. Users can record custom task demonstrations, adapt the model to new domains, and run standardized benchmarks such as WebVoyager, Online‑Mind2Web, WebTailBench, and Deepshop through the provided eval harness.

What distinguishes MolmoWeb is its end‑to‑end openness: the same repository supplies the model, training data, synthetic data generator, and a ready‑to‑use interface for real‑time navigation. This unified stack enables reproducible experiments and serves as a baseline for further development of multimodal web agents.