本文へ移動
cccskills
無料GitHub で公開

browser-use

Control the user's Chrome through the existing persistent Node REPL and the Qwen Browser SDK.

インストール方法を見る

含まれるファイル(2)

  • SKILL.md6.2 KB
  • skill.test.ts4.2 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Browser Use

Browser Use is bundled with Qwen Code and uses the generic Node REPL MCP tools and the Qwen Chrome extension. Do not install a separate Qwen extension or start a separate Browser Use MCP server.

Setup

Install and enable the Qwen Code Chrome extension in the profile the user wants to use. On macOS and Linux, first use of this SDK automatically registers the shared Native Messaging host in the user's installation directory. A browser task opts into this local setup, which can finish before the extension connects. The SDK checks the actual connection and protocol instead of reading Chrome's extension preferences.

If the extension does not connect, ask the user to open Chrome and check the extension in the intended profile. It installs from the Chrome Web Store: https://chromewebstore.google.com/detail/qwen-code/hdhmmjclhibojdddmancfgbkleahfaph

If the store reports it is not available in the user's region, they can build it from packages/chrome-extension in the Qwen Code repository (its README) and load dist/extension through chrome://extensions (Developer mode → Load unpacked). After installation or enabling it, retry the connection.

A usable Host of the same protocol is reused. A Host installed by a newer Qwen Code is never downgraded: if setup reports one, tell the user to update Qwen Code and stop. The Host files persist after Qwen exits. The user can inspect or remove them with node <skill-base>/runtime/scripts/native-host-setup.js status or uninstall; install explicitly switches the Host to this Qwen Code's copy from the next time Chrome starts it. A later Browser Use initialization can register the Host again. If required Host files cannot be read or written, report the failing action and path so the user can resolve access and retry.

Multiple Qwen sessions can use the same profile, each controlling its own tabs. A tab held by another session reports TAB_OWNERSHIP_CONFLICT. Use another tab or wait for that session to release it. The setup below binds the runtime to the default profile through browserAgent.browsers.get('chrome'). Only when the user names a specific Chrome profile, call browserAgent.browsers.list() before that line and pass the returned ID to browserAgent.browsers.get(id) instead. The runtime stays bound to the first profile it connects to; to switch profiles, call node_repl_reset and run the setup again.

If node_repl is unavailable, configure it with:

qwen mcp add --scope user node-repl npx -y @qwen-code/node-repl-mcp@0.1.7

Then tell the user to restart Qwen Code and stop. Do not start a separate Browser Use MCP server. Screenshot metadata requires @qwen-code/node-repl-mcp 0.1.7 or later, so keep this exact pin.

Qwen reports the absolute Base directory for this skill when loading this file. Use that directory as <skill-base>. The bundled SDK ships with its dependencies: import it directly, without installing packages or registering module directories, replacing the example skill base below with that absolute path. If the import reports a missing file or an incomplete runtime, stop and report an incomplete Browser Use runtime instead of installing dependencies into the workspace:

globalThis.browserAgent ??= await (
  await import('/absolute/skill/base/runtime/index.js')
).setupBrowserRuntime();
globalThis.browser ??= await browserAgent.browsers.get('chrome');
nodeRepl.write(await browser.documentation());

Create a tab or claim an exact result returned by openTabs():

globalThis.tab = await browser.tabs.new();
await tab.goto('https://example.com/');
nodeRepl.write(await tab.playwright.domSnapshot());
const candidates = await browser.user.openTabs();
nodeRepl.write(candidates);
globalThis.tab = await browser.user.claimTab(
  candidates.find((candidate) => candidate.url === 'https://example.com/'),
);

Interaction

Observe before acting and observe again after acting. Prefer semantic Playwright locators. Use tab.dom_cua.get_visible_dom() and its node_id values when snapshot refs are clearer, and use tab.cua for visual coordinate targets.

await tab.playwright.getByRole('button', { name: 'Continue' }).click();
await tab.playwright.getByLabel('Email').type('user@example.com');
await tab.dom_cua.click({ node_id: 'e4' });
await tab.dom_cua.type({ text: 'hello' });

When an action should navigate, arm the Playwright watcher around it:

await tab.playwright.expectNavigation(
  () => tab.playwright.getByRole('link', { name: 'Next' }).click(),
  { url: '**/next.html' },
);

Render screenshots with their metadata:

await nodeRepl.emitImage(await tab.screenshot());

The metadata reaches the model only through @qwen-code/node-repl-mcp 0.1.6 or later; an older server accepts the call and silently drops it. If node_repl was registered before this skill existed, check its command with qwen mcp list, and when it pins an older version, re-register it with the qwen mcp add command from Setup above and ask the user to restart Qwen Code.

Use tab.dev.logs() for bounded console diagnostics. Arm tab.playwright.waitForEvent('download' | 'filechooser') before the action that triggers it. Handle JavaScript dialogs through tab.getJsDialog(). Use a focused tab.playwright.evaluate() only when locators cannot obtain the needed data.

Use browser.user.history() only when the task needs browser history, with a focused query and bounded result count.

Finish

Treat browser.tabs.finalize({ keep }) as the final browser action of the turn. Omitted agent-created tabs close; omitted claimed tabs are released without closing. keep is the complete set for this call. Keep a user-facing result as deliverable, and keep a live page under control as handoff. Repeat a handoff in each later turn that still needs it; omitting it from the next finalize call or closing the runtime closes an agent-created tab.

await browser.tabs.finalize({
  keep: [{ tab, status: 'deliverable' }],
});

The Node kernel preserves globalThis and top-level bindings across cells. Keep cells short, await every Browser SDK promise, and use nodeRepl.write() to return text or structured data.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Reference for briefing a subagent or a fork — what to put in the prompt, what not to delegate, and a worked example. Load before writing a delegation prompt; the Agent tool's own description carries the launch rules and the background-agent rules.

日本語の概要は準備中です。原文の説明を表示しています。

QwenLM/qwen-code2.8万2026年10月10日 更新

Use after a Codex or Claude Code feature has been implemented in Qwen Code to run the selected reference agent and Qwen Code under the same scenario, capture HTTP and terminal traces, compare request bodies, tool/function schemas, outputs, and iterate until the reproduced behavior is close enough.

日本語の概要は準備中です。原文の説明を表示しています。

QwenLM/qwen-code2.8万2026年10月10日 更新

Use when reproducing an existing Codex or Claude Code feature in Qwen Code or another agent CLI by choosing a reference agent, capturing HTTP request bodies, prompts, tool/function schemas, terminal output, and then implementing the matching behavior in the target repo.

日本語の概要は準備中です。原文の説明を表示しています。

QwenLM/qwen-code2.8万2026年10月10日 更新

autofix

無料

Review and repair current local changes until they converge, or run Qwen Code Autofix issue and review workflows from GitHub Actions.

日本語の概要は準備中です。原文の説明を表示しています。

QwenLM/qwen-code2.8万2026年10月10日 更新

batch

無料

Execute batch operations on multiple files in parallel. Automatically discovers files, splits into chunks, and processes with parallel worker agents. Use `/batch` followed by operation and file pattern. For many independent single-turn transforms (translate/rewrite/extract each file into a new file) that can wait minutes to hours, you may suggest the user type `/batch-api` themselves for the half-price asynchronous Batch API — you cannot invoke it, and it is not suited to in-place edits or tasks needing tool feedback.

日本語の概要は準備中です。原文の説明を表示しています。

QwenLM/qwen-code2.8万2026年10月10日 更新

batch-api

無料

Prepare a many-file, single-turn transform (translate, rewrite, extract) as a plan and submit it to the asynchronous, half-price DashScope Batch API; results are delivered as new files hours later. Invoke explicitly with /batch-api.

日本語の概要は準備中です。原文の説明を表示しています。

QwenLM/qwen-code2.8万2026年10月10日 更新

QwenLM のスキルをすべて見る

このスキルの問題を報告する