browser-optimizer-mcp 0.1.2

Browser Optimizer MCP — reduces LLM token usage by 80-98% for browser automation

Browser Optimizer MCP

An optimization middleware layer built on top of FastMCP and Playwright. It sits between AI agents (LLMs) and browser automation frameworks to drastically reduce token usage, execution latency, and API inference costs while maintaining high accuracy for browser workflows.

Features

• Drastic Token Reduction: Reduces LLM token usage by up to 98% by filtering out non-essential HTML elements.
• Smart Element Extraction: Identifies and extracts only interactive UI components (buttons, links, inputs).
• Semantic Caching: Implements fast in-memory caching using xxhash for instant page state retrieval.
• Delta Diffing: Computes and returns only the changes between consecutive page states.
• Automated Classification: Categorizes pages dynamically based on structural heuristics.

Quick Install

pip install browser-optimizer-mcp
browser-optimizer install

The install command automatically performs the following setup:

• Checks the local Python version (requires >= 3.11).
• Installs Playwright browser binaries (chromium).
• Detects Claude Desktop and automatically configures its MCP integration.
• Detects Antigravity IDE and automatically configures its MCP integration.
• Provides manual configuration instructions for Cursor.
• Verifies the installation process end-to-end.

Once installed, initiate the server using:

browser-optimizer start

Note: Execute browser-optimizer doctor at any time to run diagnostic checks on your environment.

---

How It Works & Process Flow

1. Execution Pipeline

When an AI agent requests context from a web page, the optimizer executes the following sequence:

1. Request Intake: The AI agent invokes extract_context with a target URL.
2. Browser Navigation: The browser manager acquires a Playwright page instance and navigates to the specified URL.
3. Markup and Accessibility Capture: The extraction module retrieves the raw HTML markup and generates a semantic ARIA snapshot of the document body.
4. State Fingerprinting: The semantic cache applies an xxhash algorithm to the raw HTML to uniquely identify the page state.
   • Cache Hit: If the hash matches an existing signature, the server returns the cached context in under 1ms, bypassing DOM parsing.
   • Cache Miss: For new or modified pages, the extraction process proceeds.
5. Context Compression: The compression module filters out styling, scripts, SVGs, and boilerplate structural elements (headers, footers). It isolates interactable components such as buttons, inputs, dropdowns, and links.
6. Task Classification: A rule-based classifier evaluates the interactive elements to determine the page category (e.g., login, product search, checkout).
7. Delta Diff Calculation: The difference engine compares the current UI elements against the last observed state, returning only the added or removed components.
8. Metrics Logging: The metrics subsystem records data volume, compressed sizes, and efficiency ratios.
9. Payload Delivery: A streamlined JSON payload containing the optimized UI elements, ARIA snapshot, and page metadata is returned to the requesting agent.

2. Process Flow Diagram

sequenceDiagram
    autonumber
    actor Agent as AI Agent (LLM)
    participant Server as MCP Server (main.py)
    participant Cache as Semantic Cache (cache.py)
    participant Browser as Browser Manager (manager.py)
    participant Ext as Page Extractor (extractor.py)
    participant Comp as Context Compressor (compressor.py)
    participant Diff as Difference Engine (diff.py)

    Agent->>Server: extract_context(url)
    Server->>Browser: navigate(url)
    Browser-->>Server: page loaded
    Server->>Cache: lookup(url, current_html)
    alt Cache Hit (HTML Unchanged)
        Cache-->>Server: Return cached context
    else Cache Miss (HTML Changed / New URL)
        Server->>Ext: extract(page)
        Ext-->>Server: raw html, title, url, ARIA tree
        Server->>Comp: compress(extracted_data)
        Comp-->>Server: cleaned UI elements list, text summary
        Server->>Cache: store(url, html, compressed_context)
    end
    Server->>Diff: compute_diff(url, current_ui)
    Diff-->>Server: added & removed UI elements
    Server-->>Agent: Return optimized JSON context

---

Primary Use Cases

The Browser Optimizer MCP is designed to facilitate various AI-driven automation workflows:

1. Cost Optimization for AI Agents: Lowers token consumption on complex web pages (such as e-commerce platforms or social media feeds) by up to 98%, significantly reducing LLM API expenses.
2. Structured Data Extraction: Enables LLM-based scrapers to identify and extract content efficiently without parsing scripts, styles, or excessive HTML nesting.
3. Automated UI and End-to-End Testing: Facilitates rapid assertion checks on UI changes by leveraging delta diffing to identify structural regressions.
4. Intelligent Form Automation: Supplies clean, interactive element trees, allowing agents to populate forms, authenticate, or interact with custom components seamlessly.
5. Real-time Page Monitoring: Observes active pages for updates using delta diffing, alerting agents only when relevant interactive components change.
6. Web Search and RAG Pipelines: Functions as an efficient crawler that removes boilerplate content and provides refined context for Retrieval-Augmented Generation workflows.
7. Accessibility Audits: Exposes semantic ARIA accessibility trees, enabling automated agents to perform compliance inspections.
8. E-Commerce Monitoring: Navigates product catalogs, categorizes page types, and extracts pricing or inventory updates with minimal data transfer.
9. Dynamic Single Page Application (SPA) Support: Manages modern JavaScript-heavy frameworks by executing Playwright locally, caching states, and delivering processed components.
10. Multi-Agent Session Sharing: Acts as a standardized MCP integration layer for multi-agent systems, allowing distinct agents to share and interact with a unified browser session.

---

MCP Tools Reference

The server exposes the following tools to the Model Context Protocol environment:

Tool	Parameters	Return Type	Description
extract_context	url (string)	CompressedContext	Navigates to a URL, performs cleanup and compression, runs page classification, and returns optimized UI and ARIA trees.
page_diff	url (string)	PageDiff	Returns deltas (added/removed elements) compared to the last observed state of this URL.
execute_action	action (string), selector (optional), value (optional)	ActionResult	Executes standard interactions (click, type, select, scroll, wait, navigate) on the active page.
summarize_page	url (string)	Dict	Produces a concise text summary detailing interactive element counts and relevant text snippets.
classify_page	url (string)	ClassificationResult	Evaluates UI elements to categorize the page (e.g., login, search, survey).
wait_until_ready	url (string), timeout (optional)	ActionResult	Navigates to a page and pauses execution until browser stabilization is achieved.
cache_lookup	url (string)	Dict	Directly queries the semantic cache for a stored context entry.
get_metrics	None	Dict	Retrieves telemetry data, including bytes saved, cache hit rate, and total actions performed.

---

Architecture and Modules

The application is structured modularly within the app/ directory to ensure maintainability and separation of concerns:

• app/browser/manager.py: Manages the async Playwright browser lifecycle. Reuses page contexts to minimize startup latency and handles navigation constraints.
• app/extractor/extractor.py: Retrieves raw HTML and leverages the Playwright .aria_snapshot() API to capture accessibility structures.
• app/compressor/compressor.py: Contains DOM filtering logic. Removes superfluous tags (scripts, styles, SVGs) and generates a structured list of UI controls.
• app/classifier/classifier.py: Implements a heuristics-based scoring algorithm to categorize pages into defined states (LOGIN, SEARCH, SURVEY, CHECKOUT, PRODUCT, DASHBOARD).
• app/diff/diff.py: Compares sequential observations of a URL and constructs a delta report utilizing composite element fingerprints.
• app/cache/cache.py: Maintains an in-memory cachetools.TTLCache indexed by URL, validated via 64-bit xxhash HTML signatures.
• app/executor/executor.py: Processes standard browser interactions deterministically.
• app/schemas/schemas.py: Defines Pydantic data models to ensure strict contract compliance across all modules and tools.
• app/metrics/metrics.py: Records data size reductions, cache performance metrics, and aggregate byte savings.

---

Setup and Installation

1. Prerequisites

• Python 3.11 or higher.
• Node.js (optional, depending on external dependencies).
• Playwright system dependencies.

2. Standard Installation

# Clone the repository
git clone https://github.com/yourusername/browser-optimizer-mcp.git
cd browser-optimizer-mcp

# Create and activate virtual environment
python -m venv venv
venv\Scripts\activate     # Windows
source venv/bin/activate  # macOS/Linux

# Install requirements
pip install -r requirements.txt

# Install Playwright browser binaries
playwright install chromium

3. Environment Configuration

Create a .env file in the root directory:

LOG_LEVEL=INFO
HEADLESS=True
CACHE_ENABLED=True
CACHE_TTL=300
CACHE_MAX_SIZE=100
BROWSER_TIMEOUT=30000

---

Client Integration

1. Claude Desktop

Add the following configuration to your claude_desktop_config.json file.

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Windows Configuration (using the cmd wrapper):

{
  "mcpServers": {
    "browser-optimizer": {
      "command": "cmd",
      "args": [
        "/c",
        "C:\\path\\to\\browser-optimizer-mcp\\venv\\Scripts\\python.exe",
        "-m",
        "app.server.main"
      ],
      "env": {
        "PYTHONPATH": "C:\\path\\to\\browser-optimizer-mcp"
      }
    }
  }
}

macOS / Linux Configuration:

{
  "mcpServers": {
    "browser-optimizer": {
      "command": "/path/to/browser-optimizer-mcp/venv/bin/python",
      "args": ["-m", "app.server.main"],
      "env": {
        "PYTHONPATH": "/path/to/browser-optimizer-mcp"
      }
    }
  }
}

2. Antigravity IDE

Append this to your mcp_config.json file.

• Windows: %USERPROFILE%\.gemini\config\mcp_config.json
• macOS/Linux: ~/.gemini/config/mcp_config.json

{
  "mcpServers": {
    "browser-optimizer": {
      "command": "C:\\path\\to\\browser-optimizer-mcp\\venv\\Scripts\\python.exe",
      "args": ["-m", "app.server.main"],
      "env": {
        "PYTHONPATH": "C:\\path\\to\\browser-optimizer-mcp"
      }
    }
  }
}

3. Cursor

1. Navigate to Settings -> Features -> MCP.
2. Select + Add New MCP Server.
3. Apply the following settings:
   • Name: browser-optimizer
   • Type: command
   • Command: C:\path\to\browser-optimizer-mcp\venv\Scripts\python.exe -m app.server.main
4. Set the environment variable PYTHONPATH = C:\path\to\browser-optimizer-mcp.
5. Click Save.

---

Benchmark and Tool Comparison

A direct comparison of the Browser Optimizer MCP against standard browser automation agents demonstrates substantial efficiency improvements:

1. Performance Comparison

Metric / Feature	Standard Browser Tools	Browser Optimizer MCP
Average Token Count (Google)	~50,000+ tokens	~120 tokens (97.7% reduction)
Average Token Count (HN)	~9,000+ tokens	~1,500 tokens (87.8% reduction)
Observation Payload Type	Raw DOM or Base64 screenshots	Clean JSON UI controls + ARIA snapshot
Incremental Observations	Resends entire DOM or new screenshot	Returns only element deltas (added/removed)
Re-observation Latency	Full DOM download and parse (~1.5s)	In-memory cache lookup (~0.12ms)
Page Classification	Requires LLM API call & reasoning tokens	Instant, local rule-based heuristics (0 tokens)
Action Execution	LLM must reason step-by-step	Deterministic rule-based execution
Inference Cost	High ($0.15 - $1.00+ per step)	Extremely Low (80-95% cost reduction)

2. Running the Benchmark Suite

Execute the benchmark suite against live public pages to verify efficiency metrics locally:

$env:PYTHONPATH="."
venv/Scripts/python scripts/benchmark.py

---

Testing and Deployment

Running Unit Tests

Execute the test suite using pytest:

pytest tests/ -v

Docker Deployment

Deploy the service using Docker Compose:

docker compose -f docker/docker-compose.yml up --build

---

Contributing

Contributions are welcome. Please ensure that your pull requests adhere to the established coding standards and include appropriate unit tests. Open an issue first to discuss significant architectural changes.

---

Support and Troubleshooting

If you encounter issues during installation or runtime:

1. Run browser-optimizer doctor to diagnose environment inconsistencies.
2. Verify that your Python version is >= 3.11.
3. Check the logs/ directory for detailed exception traces.
4. Open an issue on the repository if the problem persists.

---

License

Distributed under the MIT License. See LICENSE for more information.

Copyright (c) 2026 Manthan.