dsh-plugin-browser-use

🌐 中文版 | English

Browser automation for dsh (DeepSeek Harness). Registers webclaw3 as native wc3_* tools that drive your own logged-in Chrome — skip re-login, auth popups, and most anti-bot walls.

• Scrape & export — search results, tables, downloads
• Crawl login-walled pages — 小红书, WeChat official accounts, Weibo, Taobao, admin dashboards…
• Interact — click buttons, fill forms, post content
• Reach SPA internals — React fiber / Vue data / __NEXT_DATA__

How it works

dsh (web profile)
 └─ this plugin: wc3_* tools + browsing-philosophy system prompt (auto-starts relay)
     └─ webclaw3 engine (webclaw3/, upstream fatmind/webclaw3)
         └─ relay (HTTP 127.0.0.1:3459 ↔ WebSocket)
             └─ wc3-chrome extension ──> your logged-in Chrome

The plugin is a thin adapter: it translates the relay's HTTP API into dsh tools and distills webclaw3's browsing philosophy (SKILL.md) into the system prompt. The engine and docs live upstream and are synced via scripts/sync-webclaw3.sh.

Note: in dsh you should not also install webclaw3 as a skill — this plugin already handles engine loading and usage guidance. Installing both causes duplicate instructions and conflicting relays.

Install

Prerequisites: dsh, Node >= 22, pnpm, Chrome.

curl -fsSL https://raw.githubusercontent.com/fatmind/dsh-plugin-browser-use/main/scripts/install.sh | bash

Or from a local checkout:

git clone https://github.com/fatmind/dsh-plugin-browser-use
cd dsh-plugin-browser-use && ./scripts/install.sh

The script installs dependencies → registers the plugin into the web profile (link:) → downloads & unzips the wc3-chrome extension. Then:

1. Open chrome://extensions, enable Developer mode, and Load unpacked the extension folder printed by the script;
2. Restart dsh (profiles load at startup);
3. Tell dsh "check the browser environment" — the model runs wc3_doctor and reports what's missing.

Re-running the installer updates the plugin.

Tools

Tool	relay op	Purpose
wc3_doctor	-	Channel self-check (Node/engine/relay/extension); auto-starts the relay
wc3_tabs	tab.list	List tabs and get tabIds; [当前]=active tab, [默认]=wc3's default target
wc3_tab_open	tab.create	Open a URL in a background tab (doesn't steal focus); becomes the new default
wc3_tab_close	tab.close	Close a tab you opened; clears the default target if it was closed
wc3_tab_activate	tab.list + local state	Pin a tab as the default operation target (no more "current tab" drift)
wc3_snapshot	page.ariaTree	Aria-tree snapshot with [ref_N] ids for interaction
wc3_get_text	page.getText	Full-page plain text for reconnaissance
wc3_eval	page.eval	Run arbitrary JS in the page; preferred for structured extraction
wc3_click / wc3_fill / wc3_scroll_to	page.click / page.fillForm / page.scrollTo	Interact by ref_N (React/Vue-safe)
wc3_search	page.search	Find page elements by keyword
wc3_wait_for	page.waitForElement	Wait for an element to appear

Page tools (snapshot/get_text/eval/click/fill/scroll_to/search/wait_for) accept tabId optionally — when omitted, they act on wc3's most recently active tab (newly opened, last used, or the active tab; marked [默认] in wc3_tabs). wc3_tab_close keeps tabId required to prevent accidental closes. If the extension briefly disconnects, tools auto-retry (self-heal within ~10s) before reporting an error.

Tool results are plain text with user-facing guidance, capped at 50,000 chars by default (configurable), truncated when exceeded.

Configuration

Edit packages/browser-use/cordis.patch.yml after install if needed:

Field	Default	Description
relayHost / relayPort	127.0.0.1 / 3459	Relay address
toolTimeoutMs	60000	Per-call timeout (ms)
maxResultChars	50000	Max chars per result
autoStartRelay	true	Auto-run webclaw3.mjs start when the relay is down
engineDir	"" (auto-located)	Engine directory; or set DSH_WC3_ENGINE_DIR

License

MIT