v0.3

What's New in v0.3

New Features

• PDF support in scrape_tool — extract content from PDF pages directly; specify individual pages with pages=[1,5,10]
• OAuth 2.0 + PKCE authentication — built-in OAuth flow for sites that require it
• WebMCP integration — agents can discover and call custom tools exposed by websites via the WebMCP protocol
• Loop detection — LoopGuard detects page cycles and repeated failed retries, with prompt rules to break out automatically
• keep_alive + disconnect() — keep the browser alive across agent runs and disconnect explicitly when done
• within_viewport parameter on get_state — pass within_viewport=False to get all interactive elements across the entire DOM regardless of scroll position
• Scroll position hints — browser state now includes scroll percentage and position hints for the agent

Improvements

• Unified semantic tree — DOMNode replaces separate TreeNode/TreeNodeData types; tree is now built from real DOM parent-child traversal instead of XPath reconstruction
• Richer semantic tree output — shows id/class in CSS selector notation, and role when it differs from tag
• Improved textual element detection — additional tags and correct inline text extraction
• DOM capture timing — logs state_capture_ms and screenshot_capture_ms for performance visibility
• Multiple performance optimizations across the agent loop
• Migrated to uv package manager
• Removed Playwright dependency — fully CDP-native via bundled src/cdp/ module

Bug Fixes

• Fixed PDF text extraction (switched to get_text('html') + markdownify)
• Fixed done_tool over-condensing the final output
• Fixed bounding boxes disappearing when page is scrolled
• Fixed viewport element filtering to correctly account for scroll offset
• Fixed scroll position key names in DOM viewport filtering
• Fixed sub-frame/worker crash handling in CrashWatchdog
• Fixed 10 s _wait_for_page timeouts by tracking navigation state
• Fixed browser stability and agent crash handling
• Fixed Gemini tool-calling when thought signature is absent

v0.2

Feature

• Improved the grounding to handle more corner cases.

Fix

• Fixed the bug that causes stucking in the pages of pdf or blank pages.
• Removed redundant parts in the agent implementation

v0.1

Key Features & Updates

• Dual Agent Modes: Supports both non-vision and vision-based agent operation (to support both LLM and VLM).
• Scrollable vs. Interactive Elements: A clear separation improves DOM recognition and interaction.
• Scrolling Logic: Enables scrolling through distinct webpage sections, including nested containers.
• HTML → Markdown: Upgraded to markdownify in the Scrape Tool for better content conversion.
• Tab Management: Tracks the number of open tabs, active tab, and supports basic tab control.
• Extensible Tools: Add custom tools to the agent via the additional_tools parameter.
• Iframe & Shadow DOM Access: Enhanced ability to interact with embedded or encapsulated elements.
• Structured Output: Returns well-defined BaseModel outputs using the structured_output parameter.
• Human-in-the-Loop: Add manual checkpoints in the workflow via the include_human_in_loop parameter (thanks @tanmaysk001!)
• Inference Wrapper: Fixed the bug in the open router implementation (thanks @thecoderwithHat)
• Navigation Fixes: Improved handling of edge-case navigations across complex sites.