Agent Browser

• Automating web interactions — You need an AI agent to navigate pages, click buttons, fill forms, or perform multi-step workflows on real websites.
• Extracting structured data — You want to snapshot a page's accessibility tree or DOM to pull text, attributes, or element counts from dynamic, JavaScript-rendered content.
• Filling forms programmatically — You're submitting login forms, search queries, or multi-field forms as part of an automated pipeline.
• Testing web UIs — You need lightweight, CLI-driven browser testing with screenshots, PDF export, or video recording of interactions.

• Static HTML scraping — If the target page is plain HTML with no JavaScript, a lighter HTTP client or HTML parser will be faster and simpler.
• High-concurrency crawling at scale — Agent Browser is optimized for sequential agent workflows, not distributed web crawling across thousands of URLs simultaneously.

Run agent-browser snapshot -i to get a list of all interactive elements with stable @ref identifiers (e.g. @e1, @e2). Use these refs directly in subsequent click, fill, hover, or get commands — no CSS selectors or XPath required.

Supports click, double-click, focus, fill, type, key press/hold/release, hover, checkbox check/uncheck, dropdown select, drag-and-drop, file upload, and scroll — covering virtually every user interaction a browser supports.

Save authenticated browser state (cookies, storage) to a JSON file with agent-browser state save, then reload it in future sessions to skip re-authentication. Parallel isolated sessions via --session <name> allow multiple simultaneous browser contexts.

Capture full-page screenshots, save pages as PDFs, or record browser interactions as .webm video files. Recording preserves cookies and storage from the current session, making it ideal for creating reproducible demos.

Intercept, block, or mock HTTP requests with agent-browser network route. Filter and inspect tracked requests to debug API calls or simulate offline/error states during testing.

When refs aren't available, use semantic locators to find elements by ARIA role, visible text, label, or CSS selector — for example, agent-browser find role button click --name "Submit" or agent-browser find text "Sign In" click.

AI エージェントがターゲット URL を開き、ページをスナップショットして ref でフォームフィールドを特定し、agent-browser fill で各フィールドを入力し、クリックで送信、その後ネットワークアイドルを待って再スナップショットで成功を確認します——すべて人間の介入なしで実行されます。

一度インタラクティブにログインし、agent-browser state save auth.json でセッションを保存して、以降のすべてのエージェント実行時に再読み込みします。これにより、繰り返しのログインフローを避け、自動化スクリプトから認証情報を切り離せます。

JavaScript でレンダリングされたページに移動し、agent-browser snapshot を実行してアクセシビリティツリー全体をキャプチャ

Script a sequence of browser interactions, capture screenshots or a video recording at each step, and export a PDF of the final state — giving QA pipelines visual proof of expected behavior across test runs.