blurr vs Browser Use

Blurr implements a multi-layer voice activation system combining manual tap-based triggering via DeltaSymbolView, persistent wake-word detection using Picovoice engine in EnhancedWakeWordService, and Android RoleManager integration for default assistant role. Voice input is captured, transcribed via speech-to-text, and routed to the conversational agent service which interprets natural language intent and triggers the AI agent execution framework. The system maintains always-on listening capability without requiring explicit app focus.

Unique: Combines Picovoice on-device wake-word detection with Android Accessibility Service for full-system UI automation, avoiding cloud-dependent voice processing while maintaining always-on listening without explicit app activation

vs alternatives: Unlike cloud-based voice assistants (Google Assistant, Alexa), Blurr processes wake words locally for privacy and offline capability, while unlike browser automation tools, it operates at the Android OS level with native accessibility APIs for true cross-app automation

Blurr's perception layer leverages Android's AccessibilityService to read the complete UI hierarchy (AccessibilityNodeInfo tree) from the currently visible screen, extracting semantic information about interactive elements, text content, and layout structure. This accessibility tree is serialized into a structured representation that the LLM can reason about, enabling the agent to understand which buttons, text fields, and interactive components are available without relying on image recognition or OCR. The system captures both the visual state and the semantic meaning of UI elements.

Unique: Uses Android AccessibilityService for semantic UI tree extraction rather than vision-based screen analysis, providing structured element information without image processing overhead while respecting app security boundaries

vs alternatives: More reliable than vision-based UI detection (which fails with dynamic content) and faster than OCR-based approaches, but requires accessibility permission and cannot penetrate apps that block accessibility tree access

Blurr integrates Firebase Analytics to track user behavior, task execution patterns, and feature usage. Firebase Crashlytics captures runtime errors and exceptions, providing crash reports and stack traces for debugging. The system logs key events (task execution, permission grants, subscription changes) to Firebase for analytics. This data enables the developers to understand user behavior, identify bugs, and optimize the product. Firebase also provides real-time dashboards for monitoring app health and user engagement.

Unique: Integrates Firebase Analytics and Crashlytics to provide real-time usage tracking, crash monitoring, and user behavior insights, enabling data-driven product optimization and debugging

vs alternatives: More comprehensive than simple error logging (includes user behavior analytics and real-time dashboards), but adds network overhead and privacy considerations

Blurr stores user data locally using Android's persistence mechanisms (likely SharedPreferences for simple data, Room database for complex data structures). Sensitive information (API keys, authentication tokens, user preferences) is encrypted using Android's EncryptedSharedPreferences or similar encryption libraries. The system manages data lifecycle (creation, update, deletion) and handles data migration across app versions. Local storage enables offline operation for certain features and reduces dependency on cloud services.

Unique: Implements encrypted local storage using EncryptedSharedPreferences and Room database, providing secure persistence of sensitive data while maintaining offline capability and reducing cloud dependency

vs alternatives: More secure than unencrypted local storage but less convenient than cloud sync; requires careful key management and is vulnerable to device compromise

Blurr enables automation workflows that span multiple applications, maintaining context and state as the agent navigates between apps. The system detects app transitions (via AccessibilityService), preserves task context across app boundaries, and adapts the UI perception and action execution to each app's specific interface. This allows complex workflows like 'open email, find message from John, extract phone number, open contacts, add new contact with that number' where the agent must understand context across three different apps. The agent maintains a unified task model that abstracts away app-specific details.

Unique: Implements cross-app workflow orchestration with unified task modeling and context preservation, allowing the agent to maintain state and task progress as it navigates between multiple applications with different UI patterns

vs alternatives: More sophisticated than single-app automation (handles complex multi-app workflows) but more fragile than app-specific automation (requires careful context management and app-specific handling)

Blurr implements robust error handling that detects when actions fail (element not found, action timed out, unexpected UI state) and attempts recovery. The system includes fallback strategies: retry with adjusted timing, alternative action paths (e.g., using menu instead of direct button), and user escalation (asking user for help or manual intervention). Error detection works by comparing expected UI state (from LLM reasoning) with actual UI state (from accessibility tree) after each action. The system logs errors for debugging and learns from failures to improve future action selection.

Unique: Implements multi-level error recovery with fallback strategies, retry logic, and user escalation, detecting action failures by comparing expected vs actual UI state and attempting recovery before giving up

vs alternatives: More robust than simple retry logic (includes fallback strategies and escalation) but more complex to implement and debug than stateless error handling

Blurr integrates Google Gemini API as the reasoning engine that receives the current screen state (accessibility tree), user intent (voice command), and task history, then generates the next action to execute. The LLM operates in an agentic loop: it analyzes the current UI state, reasons about the user's goal, selects the most appropriate action (tap, scroll, type, etc.), and provides structured action output that the execution layer interprets. The system maintains conversation context across multiple turns, allowing the agent to handle multi-step workflows that require understanding previous actions and adapting to screen changes.

Unique: Implements a closed-loop agentic architecture where Gemini LLM receives structured accessibility tree data and generates typed action outputs that directly map to Android UI automation APIs, with explicit error recovery and context management for multi-step workflows

vs alternatives: More sophisticated than rule-based automation (handles dynamic UIs and novel scenarios) and more reliable than vision-based agents (semantic tree data is more stable), but requires API access and introduces latency compared to local models

Blurr's action execution layer translates LLM-generated action specifications into native Android UI automation commands via the AccessibilityService API. The system supports multiple interaction primitives: single/multi-touch taps at specific coordinates, swipe/scroll gestures with configurable velocity and direction, text input via keyboard simulation, and long-press interactions. Actions are queued and executed sequentially with timing controls to allow UI animations to complete between actions. The execution layer includes error detection (checking if expected UI changes occurred after an action) and fallback mechanisms for failed interactions.

Unique: Implements a queued, error-aware action execution system that translates high-level action specifications into AccessibilityService API calls with built-in timing controls, error detection, and fallback mechanisms for handling UI animation delays and interaction failures

vs alternatives: More reliable than coordinate-based image automation (uses semantic element information) and more flexible than simple tap/swipe APIs (supports complex gesture sequences and error recovery), but requires AccessibilityService permission and cannot bypass app-level security restrictions

browser-use/browser-use | DeepWiki Loading... Index your code with Devin DeepWiki DeepWiki browser-use/browser-use Index your code with Devin Edit Wiki Share Loading... Last indexed: 17 May 2026 ( 933e28 ) Overview System Architecture Installation and Setup Quick Start Examples Agent System Agent Core and Execution Loop Message Manager and Prompt Construction Agent State and History Management System Prompts and Output Formats Skills Integration Agent Configuration and Settings Loop Detection and Behavioral Nudges Message Compaction System Memory and Follow-up Tasks Judge System and Trace Evaluation Browser Session Management BrowserSession Lifecycle Browser Profile Configuration SessionManager and CDP Session Pool Target and Frame Management Navigation and Tab Control Event-Driven Architecture Event System Overview Event Types Reference Watchdog Pattern and Base Classes Core Watchdog Implementations DOM Processing Engine DOM Tree Construction DOM Serialization Pipeline Interactive Element Detection Visibility Calculation and Coordinate Transformation Screenshot Highlighting System Browser State Summary Markdown Extraction and HTML Serialization Tools and Action System Tools Registry and Action Models Built-in Actions Reference Action Execution Pipeline Custom Tools and Extensions Click Action Deep Dive Input Action and Autocomplete Detection FileSystem Integration Br

System Architecture | browser-use/browser-use | DeepWiki Loading... Index your code with Devin DeepWiki DeepWiki browser-use/browser-use Index your code with Devin Edit Wiki Share Loading... Last indexed: 17 May 2026 ( 933e28 ) Overview System Architecture Installation and Setup Quick Start Examples Agent System Agent Core and Execution Loop Message Manager and Prompt Construction Agent State and History Management System Prompts and Output Formats Skills Integration Agent Configuration and Settings Loop Detection and Behavioral Nudges Message Compaction System Memory and Follow-up Tasks Judge System and Trace Evaluation Browser Session Management BrowserSession Lifecycle Browser Profile Configuration SessionManager and CDP Session Pool Target and Frame Management Navigation and Tab Control Event-Driven Architecture Event System Overview Event Types Reference Watchdog Pattern and Base Classes Core Watchdog Implementations DOM Processing Engine DOM Tree Construction DOM Serialization Pipeline Interactive Element Detection Visibility Calculation and Coordinate Transformation Screenshot Highlighting System Browser State Summary Markdown Extraction and HTML Serialization Tools and Action System Tools Registry and Action Models Built-in Actions Reference Action Execution Pipeline Custom Tools and Extensions Click Action Deep Dive Input Action and Autocomplete Detection FileS

Agent System | browser-use/browser-use | DeepWiki Loading... Index your code with Devin DeepWiki DeepWiki browser-use/browser-use Index your code with Devin Edit Wiki Share Loading... Last indexed: 17 May 2026 ( 933e28 ) Overview System Architecture Installation and Setup Quick Start Examples Agent System Agent Core and Execution Loop Message Manager and Prompt Construction Agent State and History Management System Prompts and Output Formats Skills Integration Agent Configuration and Settings Loop Detection and Behavioral Nudges Message Compaction System Memory and Follow-up Tasks Judge System and Trace Evaluation Browser Session Management BrowserSession Lifecycle Browser Profile Configuration SessionManager and CDP Session Pool Target and Frame Management Navigation and Tab Control Event-Driven Architecture Event System Overview Event Types Reference Watchdog Pattern and Base Classes Core Watchdog Implementations DOM Processing Engine DOM Tree Construction DOM Serialization Pipeline Interactive Element Detection Visibility Calculation and Coordinate Transformation Screenshot Highlighting System Browser State Summary Markdown Extraction and HTML Serialization Tools and Action System Tools Registry and Action Models Built-in Actions Reference Action Execution Pipeline Custom Tools and Extensions Click Action Deep Dive Input Action and Autocomplete Detection FileSystem I

browser-use/browser-use | DeepWiki Loading... Index your code with Devin DeepWiki DeepWiki browser-use/browser-use Index your code with Devin Edit Wiki Share Loading... Last indexed: 17 May 2026 ( 933e28 ) Overview System Architecture Installation and Setup Quick Start Examples Agent System Agent Core and Execution Loop Message Manager and Prompt Construction Agent State and History Management System Prompts and Output Formats Skills Integration Agent Configuration and Settings Loop Detection and Behavioral Nudges Message Compaction System Memory and Follow-up Tasks Judge System and Trace Evaluation Browser Session Management BrowserSession Lifecycle Browser Profile Configuration SessionManager and CDP Session Pool Target and Frame Management Navigation and Tab Control Event-Driven Architecture Event System Overview Event Types Reference Watchdog Pattern and Base Classes Core Watchdog Implementations DOM Processing Engine DOM Tree Construction DOM Serialization Pipeline Interactive Element Detection Visibility Calculation and Coordinate Transformation Screenshot Highlighting System Browser Sta