dsh-playwright

LLMs are great at reasoning and generating code, but they rarely get to "see" a live browser page. dsh-playwright bridges that gap for DeepSeek Harness: through Chrome DevTools Protocol and Playwright, you and the model share a single isolated Playwright page. You watch repaint frames stream into the Browser Use canvas while the model reads semantic snapshots, moves the pointer, clicks, drags, types, and even returns a PNG screenshot to a multimodal model for verification.

The toolset covers nearly every day-to-day web interaction: opening and reloading HTTP(S) URLs, extracting the title and visible text with temporary element IDs, targeting by ID or screenshot coordinates with left, right, or middle button down-move-up sequences for clicks and drags, keyboard typing, single-key and chord presses (Control+A, Meta+K), and game-style hold-and-release. Every DSH session owns its own BrowserContext and Page so parallel conversations never cross-contaminate; if Chrome exits unexpectedly the next operation relaunches it and restores the affected session to its last top-level URL. By default the plugin blocks localhost and private-network targets, and its RPC and screencast endpoints accept loopback same-origin clients only.

This is ideal for developers who want AI-assisted web interaction, UI verification, or debugging: watch the pointer glide over controls in the panel while the model acts on semantic IDs or screenshot coordinates, then confirm the result with a click-marked screenshot—all without leaving the chat window.