dsh-browser-tool

DeepSeek Harness agents are strong at text reasoning but lack a direct window into what renders on a live web page. dsh-browser-tool bridges that gap by letting the model open a Chromium tab, execute JavaScript, interact with the DOM, capture a screenshot, and receive both the text stream and the image block back inside the conversation, completing a full observe-act-observe loop within a single agent session.

The API is shaped around a clean open / run / close contract. open provisions a browser through four interchangeable sources: an in-process headless Chromium with optional stealth injection, a detached desktop application attached over its idle CDP port, a pre-existing CDP endpoint, or a local Browser Relay that takes over the user's running Chrome session with tab-group queuing and debugger-ban recovery. run executes an async function body inside the tab, exposing the full Tab API (goto, observe, screenshot, click, type, fill, press, scroll, drag, waitFor, evaluate, uploadFile, and more). Returned screenshots are resized to the model pixel budget via sharp before arriving as image content blocks. Chromium binary resolution walks a priority list that prefers system installs and Chrome for Testing builds already cached by other tools, falling back to a download only when every other source is absent.

The plugin suits developers who need web-level interaction inside an agent workflow: automated information extraction, end-to-end UI verification, competitive-page inspection, multi-step form filling, and similar tasks. It ships as a self-contained npm package requiring no local build, mounts as a DSH profile bundle, and is released under the MIT license.