Browser Use changelog: what's new each month
Every stable Browser Use release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 12 months covered; the current month updates daily.
November 2026 · so far
8 releases: 1.0.0 → 1.0.7
November brought Browser Use to version 1.0, a major release that stabilizes the API, removes legacy code paths, and adds server-side browser automation. The month also introduced an AI judge to validate task completion and improved cloud browser security and isolation.
- All 0.x imports raise ImportError naming the replacement and linking MIGRATING.md.
- Browser() now defaults to cloud browser; local tasks need Browser('chromium'); Browser('chrome') attaches to your own Chrome only.
- Browser(local=True) is now an alias for Browser('chromium').
- Browser('chrome') no longer attaches to arbitrary browsers answering on 127.0.0.1:9222.
Highlights
- Version 1.0 is a breaking change: all 0.x imports now raise ImportError with migration guidance, and configuration options like include_in_memory are deprecated.
- MCP and CLI now run their own Chromium on servers and containers, with clear error messages when no browser is available.
- An AI judge automatically evaluates whether each run succeeded by checking the trace, screenshots, and final page against the task.
- BU 4.0 is now the default model, delivering higher quality results at increased cost.
- Browser attachment is more selective: Chrome only attaches to browsers with live DevToolsActivePort in your profile, and cloud browsers are the default.
New
- AI judge validates run success with judgement field containing success status, failure_reason, description, and prompt_injection_detected.
- MCP and CLI automatically drive their own Chromium on servers, containers, and CI boxes, installed on first use.
- Browser-use feedback system: users and agents can report issues via CLI, MCP, and report_issue() calls.
- CLI runs its own Chromium when no browser has remote debugging enabled or there is no display, persisting across calls.
- BU_NAME environment variable lets callers get their own Chromium when Chrome cannot be used.
- Long browser_exec results in MCP now reply within 50 seconds with output so far, preventing timeouts for clients with 60-second limits.
- Agent judge observes the trace and final page to ground gotchas in what the run actually observed.
- DevToolsActivePort detection for Chrome in another process namespace, enabling containers and Cursor on Linux.
Improved
- run() no longer raises on configuration errors or browser startup failures; rejected keys, models, and browser failures return status='error'.
- max_steps default reduced from 500 to 100, where one step is one model request.
- CDP calls from CLI and MCP wait up to 20 seconds like the agent; frozen pages fail fast with dialog handling guidance.
- Screenshots are dropped in one batch once 20 tool results carry one, reducing memory overhead.
- Cancelled cloud runs copy downloads within 10 seconds instead of holding the browser up to 300 seconds.
- BU_CDP_URL failures now fail in 3 seconds with a message naming the value and both solutions.
- Windows: piped CLI code and MCP output now use UTF-8 encoding; open() in MCP browser_exec code reads and writes UTF-8.
- MCP with allowed_domains can no longer be bypassed through a cloud browser.
Fixed
- MCP exits cleanly when the client closes the pipe before the reply, instead of exit 1 with ExceptionGroup.
- Read-only HOME directory errors now name the environment variable to set.
- timeout parameter now ensures the final answer arrives before the deadline, and provider retries respect it.
- ActionResult warns about every ignored 0.x field, not only the first; include_in_memory is ignored and all tool results stay in conversation.
- Feature detection: hasattr(browser_use, '<removed 0.x name>') returns False.
- Each run gets its own browser-harness workspace for isolation.
October 2026
4 releases: 0.13.11 → 0.13.14
October brought improved OpenAI integration, better support for Browser Use Cloud, and fixes for display scaling and sensitive data handling. The month focused on strengthening cloud capabilities and cross-platform compatibility.
Highlights
- New OpenAI integration with browser tools support and simplified imports from browser_use.integrations.openai
- Fixed screenshot and interaction alignment on scaled displays so clicks and crops match what users see
- Improved handling of Browser Use Cloud tabs and browser recovery with clearer error messages
- Enhanced sensitive data redaction in saved history for better security
- Support for creating small binary images directly for upload workflows
New
- OpenAI Python 3.31+ integration available via browser-use[openai-tools] package
- 18 default browser tools and six optional actions through Browser Use Actor
- Ability to create tiny PNG, GIF, JPG, and WebP images for upload flows
- Support for borrowed browser sessions alongside local browsers and Browser Use Cloud
Improved
- Page observations, form state detection, file access, and browser recovery
- Handling of Cloud tab URLs during navigation to prevent temporary empty states
- Error messages for missing Cloud browser keys are now more helpful
- URL and timeout settings now preserved across OpenAI 2.x and 3.x model wrappers
- Screenshot, crop, and pointer coordinate alignment on displays with scaling
- Browser Use Cloud SDK V4 import support
Fixed
- Gemini requests no longer drop system messages
- Sensitive data nested below list boundaries is now properly redacted in saved history
- Data URL scheme and media type handling is now case-insensitive
- DOM hidden code style matching is now case-insensitive
- Temporary empty Cloud tab URLs during navigation
- Failed browser reconnect initialization cleanup
September 2026
2 releases: 0.13.9 → 0.13.10
September brought dependency updates, model defaults adjustments, and improvements to error reporting. The month focused on stability and compatibility with the latest tool frameworks.
Highlights
- Browser Harness upgraded to 0.1.13 with improved compatibility.
- Migrated to MCP Python SDK 2.1.1 for better model context protocol support.
- All runtime dependencies are now exactly pinned for reproducible builds.
- Unknown MCP tool calls now properly reported as application errors instead of success.
New
- Pydantic-settings 2.15.0 added as explicit runtime dependency.
Improved
- DeepSeek default model updated to deepseek-v4-flash.
- ChatBrowserUse bu-2-0 restored as default, with mini available as opt-in.
- PDF generation now properly escapes plain text content.
- File replace operations report missing target text in error messages.
- Dependency versions bumped for security and compatibility.
Fixed
- Resolved three pypdf Dependabot security advisories by upgrading to version 6.16.2.
- Fixed missing target text reporting in filesystem replace operations.
- Corrected plain text escaping in generated PDFs.
August 2026
1 release: 0.13.8
August brought reliability improvements and model updates to Browser Use, with fixes for edge cases in agent execution and DOM handling.
Highlights
- Cerebras model IDs and defaults updated
- Fixed agent crashes when handling empty action lists or zero division scenarios
- Improved DOM chunk handling for content with trailing newlines
- Better recovery of Anthropic tool arguments when models serialize calls as text
Improved
- Agent now properly recovers Anthropic tool arguments when the model returns them as text
- Temporary user data directory creation now returns Path objects consistently
- Agent file writes now use UTF-8 encoding
Fixed
- Agent no longer crashes when lastaction() is called with an empty action list
- Budget warning injection no longer causes ZeroDivisionError
- DOM final chunk character end position now clamped correctly for content with trailing newlines
July 2026
5 releases: 0.13.3 → 0.13.7
July brought the launch of Browser Use CLI 3.0 alongside numerous stability improvements and fixes for navigation, error handling, and extraction reliability.
Highlights
- Browser Use CLI 3.0 powered by Browser Harness is now available.
- Browser Use skill can now be installed directly by coding agents.
- Improved navigation readiness detection fixes timeout and same-document navigation issues.
- LLM output truncation is now detected and can fall back to alternative models instead of failing.
- Skill installation behaviour improved across Claude Code, Codex, Cursor, Gemini, and other agent platforms.
New
- MCP server for CLI 3.0 with shared MCP runner and maxdim guard.
- Support for MCP registry.
- Accept bu-qa-1 model alias in ChatBrowserUse.
- Shim for legacy CLI commands with hints for unknown commands.
Improved
- Navigation readiness detection now uses per-target lifecycle event storage.
- Markdown extraction no longer destroys URLs or drops long link lines.
- Plain BrowserErrors remain recoverable and structured error memory is preserved through executeaction.
- Visibility check no longer mutates shared snapshot bounds.
- CLI startup is now faster.
- Skeleton-screen loading hint is gated on actual in-flight network requests.
- Selector indices are kept unique across CDP sessions.
- File extraction handles file:// URLs correctly.
Fixed
- Navigation readiness timeout handling and same-document navigation support.
- Markdown extraction destroying URLs and dropping long link lines.
- LLM output truncation detection and fallback to alternative models.
- Clear-to-empty on React controlled inputs using native value setter.
- Nested cross-origin iframe extraction filtering.
- Tool name in flash prompts corrected from replacefilestr to replacefile.
June 2026
3 releases: 0.13.0 → 0.13.2
June brought compatibility updates with the latest Claude models and improved model handling across providers. The release cycle included stability fixes and preparation of core infrastructure for broader SDK support.
Highlights
- Added support for Claude Fable 5 model
- Improved model selection to accept provider-prefixed model names
- Added BU3 browser use models to Codex
- Fixed tool environment variable handling in packaged agents
- Enhanced prerelease version upgrade checking
New
- X402 skill added
- Support for Claude Fable 5
- BU3 browser use models now available
- Browser-use-core binary discovery capability
Improved
- Tool choice handling for Claude Fable 5 with automatic mode when thinking is enabled
- Model selection now accepts provider-prefixed model names for more flexible configuration
- Rust SDK server validation and cleanup
- Beta SDK browser resources management
Fixed
- Prerelease version upgrade check
- Packaged agent tools environment variables
- PyPI publishing now gated behind release environment verification
May 2026
3 releases: 0.12.7 → 0.12.9
May brought security patches, daemon improvements, and better defaults for LLM model selection. The month focused on fixing vulnerabilities and hardening the browser use agent.
Highlights
- Security fixes for aiohttp memory exhaustion and requests path-traversal vulnerabilities
- Unix socket now restricted to owner-only access for improved daemon security
- Per-session authentication tokens added to daemon connections
- Default LLM model changed to bu-2-0 for better out-of-the-box performance
- CLI received another major update with multiple correctness fixes
New
- Client header added to GoogleChat integration
- OpenRouter pricing fallback for token cost calculation
- Session ID now passed to judge LLM calls for better context
Improved
- Model documentation improved
- HistoryItem now frozen to prevent accidental modifications
- LLM ainvoke patching uses setattr to avoid Pydantic crashes
- Gemini 3 Flash Preview recommended in examples and tests
Fixed
- Security and correctness issues from code review
- KeyError in loadfromdict and IndexError in finalresult
- Evaluate tool now refuses to run on restricted browser profiles
- Screenshots no longer captured on new tab pages
April 2026
1 release: 0.12.6
April brought stability improvements and configuration enhancements, focusing on model compatibility and connection handling. The release addressed several edge cases in structured output, shell configuration, and daemon management.
Highlights
- Fixed daemon orphaning issues when connecting to external browsers via CDP
- Improved Bedrock structured output handling with schema flattening
- Added CDP navigation timeout configuration for better control over page load behavior
New
- Option to disable SignalHandler for better integration with host application signal control
- Ability to override model pricing URL for token calculation
Improved
- Default temperature set to 1.0 for Gemini-3 models for better compatibility
- Login shell usage in install script for proper PATH configuration
- Anthropic action field serialization to prevent double-encoding
Fixed
- Daemon orphaning when external browser dies on --connect or --cdp-url
- Bedrock structured output schema optimization
- Shell PATH configuration in installation
- Anthropic action field double-serialization
March 2026
5 releases: 0.12.1 → 0.12.5
March brought stability improvements and hallucination prevention across the board, with better model support and more reliable DOM handling. The month focused on data grounding, browser state accuracy, and reducing false positives in extraction tasks.
- LiteLLM is no longer installed by default; install separately with pip install litellm if needed
Highlights
- Prevent duplicate items in multi-page extraction tasks with duplicate detection
- Add strict data grounding rules to reduce AI hallucination during navigation and actions
- Exclude password fields from DOM snapshots sent to the model for security
- Enhance ChatVercel model support with improved API key handling
- Detect empty DOM after navigation and retry before reporting failure
New
- LiteLLM integration for broader model provider support
- Hard constraints checking system
- Better data grounding in the done action
- ChatLiteLLM wrapper for using additional model providers
Improved
- Enhanced browser state format in system prompts
- Improved multi-page data extraction to avoid duplicates
- Better detection and handling of empty DOM states after navigation
- Refined CDP request handling for localhost connections
- Updated ChatVercel model support and API key handling
- Improved open-source model prompt guidance in documentation
Fixed
- Proxy authentication handler no longer overwrites tab tracking
- Storage state restoration now properly handles cookies and localStorage
- Prevent removeHighlights() from blocking screenshot operations
- Fix browser state format issues in system prompts
- Prevent compaction hallucination in extraction tasks
- Corrected uv installation detection for curl-installed versions
February 2026
10 releases: 0.11.6 → 0.12.0
February brought significant stability improvements and new capabilities to Browser Use, with better handling of edge cases, enhanced authentication support, and new tools for content extraction and interaction. The agent system saw multiple refinements to handle complex web scenarios more reliably.
- Removed simple judge from the system.
Highlights
- Added automatic CAPTCHA solver handling via CAPTCHA watchdog to reduce manual intervention.
- New readlongcontent action for handling long webpages and files more effectively.
- Improved authentication with new Browser.fromsystemchrome() method and better real browser setup documentation.
- Fixed critical issues with browser lifecycle management, logging, and sensitive data handling.
- Added save-as-PDF export capability for agent actions via CDP.
New
- Grep tool added to agent for text searching.
- Extraction schema support with schema enforcement in extract tool.
- Basic planning capability added to agents.
- Action loop detection to prevent infinite action cycles.
- readlongcontent action for processing long webpages and files.
- Automatic CAPTCHA solver handling via CAPTCHA watchdog.
- Browser.fromsystemchrome() method for using system Chrome instances.
- Save-as-PDF agent action for CDP-based page export.
Improved
- Increased LLM timeout to 75 seconds for Gemini and Browser Use models.
- Increased default step timeout from 120 to 180 seconds.
- Increased agent memory length from 1000 to 10000 characters.
- Improved filename sanitization.
- Enhanced multiact functionality for agents.
- Improved Markdown chunker processing.
- Better text field handling and autocomplete field support.
- WebSocket reconnection support for remote browser CDP connections.
Fixed
- Fixed agent eventbus clearing during session cleanup.
- Fixed browser closure detection to stop agents appropriately.
- Fixed consecutive failures counter resetting on multi-action steps.
- Fixed radio button click occlusion when label overlaps input.
- Fixed sensitive data replacement when LLM omits secret tags.
- Fixed download monitoring for fetch requests and blob downloads.
January 2026
3 releases: 0.11.3 → 0.11.5
January brought stability improvements to element matching and rerun functionality, along with new model support and CLI enhancements. The month focused on better reliability for complex browser interactions and expanded integration options.
Highlights
- Added support for new browser use models including Gemini 3 Flash Preview and Browser Use 2-0.
- Introduced Browser Use CLI with skill support for direct command-line interaction.
- Improved element rerun stability through cascading element matching and unique attribute detection.
- Added ability to detect JavaScript event listeners on clickable elements.
- Extended context windows for supported models with 4096-token prompts for Opus and Haiku 4.5.
New
- Gemini 3 Flash Preview model support.
- Browser Use 2-0 compatibility.
- Browser Use CLI with skill support.
- JavaScript event listener detection for clickable elements.
- BROWSERUSEDISABLEEXTENSIONS environment variable to disable browser extensions.
- Thinking level configuration for Gemini 3 Pro model.
- ChatBrowserUse support for BROWSERUSEAPIKEY environment variable.
- Unique attribute matching for improved element identification.
Improved
- Cascading element matching for more stable element rerun identification.
- Enhanced error logging for rerun scenarios.
- Better placeholder handling for 2FA secrets in registry service.
- Character limit increased for extraction operations.
- Glob pattern matching for Playwright browser paths.
- Default thinking set to 0 for Gemini 3 models to optimize performance.
- Custom tool registry documentation expanded.
Fixed
- Element failures no longer skipped by default in rerun history.
- Improved axname inclusion for rerun history element hashing.
- Fixed Playwright browser path detection with corrected glob patterns.
- Resolved JavaScript event listener detection for clickable elements.
December 2025
3 releases: 0.11.0 → 0.11.2
December brought cloud support and expanded LLM provider options to Browser Use, along with performance improvements and fixes for output handling and browser tab management.
Highlights
- Added cloud support for Browser Use
- Integrated Mistral chat provider with schema sanitization
- Added fallback LLM support for redundancy
- Improved startup latency with lazy imports for OpenAI and ReportLab
- Enhanced security by filtering sensitive data from history
New
- Mistral chat provider integration
- Fallback LLM support
- Optional coordinate availability in actions
- Cloud platform support
Improved
- Startup performance optimized with lazy imports
- System prompts updated
- XAI vision check refined to target specific Grok models
- Sensitive data filtering in execution history
- Structured output handling for sandboxes
Fixed
- Short line removal bug
- Supported model documentation link
- ChatOpenAI import path in documentation
- Unnecessary blank tab creation when closing background tabs
- Structured output generation for sandbox environments
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Browser Use: pricing, features and alternatives · All changelogs