Browser automation that skips screenshots and headless mode

What it does

dsh-browser is a Chrome and Firefox extension plus a bridge plugin that connects DeepSeek Harness to the browser tab you already have open. The model reads a structured text snapshot of the page—complete with a numbered inventory of buttons, links, and form fields—then clicks, types, scrolls, and navigates by referring to those numbers. A side panel provides the chat interface, and everything runs in your real profile so sessions and cookies stay intact.

The interesting bit

Instead of feeding screenshots to a vision model, the extension translates the DOM into a text-only interface with stable IDs and delta updates; the model operates the page by number, as if working from a numbered parts catalog. Multimodal images still reach the model, but only through the separate chat attachment path—never via browser screenshots.

Key highlights

• Text-first instrumentation: pages become numbered structured text, so no vision capabilities are required for browser tools.
• Privacy guardrails: passwords and payment-card values are masked and never leave the page; the assistant binds to one tab and pauses if you switch away.
• Quote-by-highlight: select text on the page and the side panel quotes it, so you can ask “explain this” without a page tour.
• Benchmark claim: in a 60-run test against a Playwright baseline, the README reports 30/30 success with roughly 20% lower latency and fewer tool calls.
• Security boundaries: bearer-token auth, loopback-only privileged gateway methods, and untrusted-content wrapping for page text.

Caveats

• Not distributed through browser stores or npm; the README warns that the unscoped dsh-browser npm package is unrelated.
• Firefox requires manual bearer-token entry because moz-extension:// UUIDs are per-install.
• Protected browser pages and extension stores are unsupported.

Verdict

Worth a look if you are already running DeepSeek Harness and want browser automation without spinning up headless Chromium or paying for vision tokens. Skip it if you need a one-click store install or fully automatic Firefox support.