Agent Browser

• You need to automate multi-step browser workflows where an AI agent must navigate, read state, and interact across multiple pages.
• You require deterministic element selection — ref-based targeting eliminates the fragility of XPath or pixel-based clicks.
• You're working with complex Single Page Applications (SPAs) where the DOM changes dynamically and snapshots need to be re-taken after each action.
• You need session isolation — running multiple simultaneous browser contexts (e.g., admin and user) without interference.

• You need screenshots or PDFs for visual analysis — use a built-in browser tool that specializes in visual inspection.
• You need browser extension integration — agent-browser does not support loading extensions.
• Your task is purely visual and does not involve interacting with page elements.

Instead of fragile CSS selectors or coordinates, agent-browser captures the accessibility tree and assigns stable @ref IDs (e.g., @e2) to each element. AI agents parse these refs from JSON output and use them for all subsequent interactions, ensuring reliable targeting even across dynamic page updates.

Supports click, fill, type, hover, check/uncheck, select, drag, scroll, press, and more — all ref-driven. Also provides get and is commands to extract text, HTML, attributes, and check visibility or enabled state.

Run multiple named sessions simultaneously with --session <name>. Each session maintains its own browser context, enabling scenarios like testing admin and user roles side-by-side without cookie or state bleed.

Save cookies and local storage to a JSON file with state save, then reload it later with state load to skip login flows entirely — ideal for repeated automation runs.

Block unwanted requests (e.g., ads), mock API responses, and inspect network activity — all from the CLI. Useful for testing and scraping without external interference.

Wait for elements, text content, URL patterns, network idle, or arbitrary JavaScript conditions. Ensures stable snapshots before the agent proceeds to the next step.

An AI agent navigates to a search engine, fills in a query using a ref-identified input, waits for results, then extracts text and links from result elements — all without screenshots or manual selector tuning.

Two simultaneous sessions — one authenticated as admin, one as a regular user — interact with the same app concurrently. Each session loads its own saved auth state, enabling parallel workflow validation.

An agent logs in once, saves the session state to auth.json, and on subsequent runs loads that state to skip login — then proceeds directly to filling and submitting forms.

For complex React or Vue apps, the agent re-snapshots after every navigation or state change, always working from a fresh accessibility tree to ensure refs are current before interacting.