dsh-browser

Why we included this project

DeepSeek Harness users who want the model to actually operate their browser will find this extension handles the wiring. It turns the active tab into structured text with a numbered inventory of interactive elements, so the agent can click, type, scroll, and move between pages without ever analyzing a screenshot. Everything happens as text, which means models without vision support work fine, and your logins, session, and cookies carry over. A side panel keeps the conversation visible while the agent works in the tab you already have open. This is a focused bridge, not a full automation platform, so it suits teams already on dsh that want to hand real browser control to their agent. A one-line installer covers macOS, Linux, and Windows, and the README benchmark shows how it finishes the same job in fewer round trips than the alternative.