dsh-browser-vision

In the realm of web automation, general-purpose models are often prohibitively expensive due to high visual processing costs. dsh-browser-vision is designed to solve this exact problem. By leveraging DeepSeek's deepseek-v4-flash-vision-exp model, it transforms the traditionally premium "seeing" capability into a cheap and fast default option. With screenshots costing at most 384 input tokens per step, this plugin reduces overall costs by two-thirds while ensuring agents can read text, charts, and non-DOM elements, fundamentally changing the economics of vision-based agents.

Beyond basic navigation, clicking, and form filling, the tool excels at structured data extraction. Whether dealing with invoices embedded in images, CAPTCHAs, or data hidden in charts, the plugin uses a JSON Schema validator to return clean, validated JSON outputs, eliminating the need for downstream parsing. Its intelligent "auto" vision mode determines whether a screenshot is necessary, optimizing resource usage without sacrificing accuracy for complex UI tasks.

This plugin is ideal for cost-conscious DeepSeek Harness users, CI/CD pipeline developers, and various Agents. If you need to run large-scale web testing and automation tasks locally or in the cloud while strictly adhering to budget constraints, dsh-browser-vision offers the most cost-effective solution on the market. It makes high-precision, low-cost web automation accessible to everyone.