Browser Use for AI Agents

True Agent Multiplexing: Run Cursor, Claude, Codex, Gemini, OpenCode, and any other custom terminal agents at the same time. The server assigns unique Agent Session IDs (ASID) to dynamically route commands without cross-session contamination. **Other extensions make you pay for this feature.**

Strict Tab Locking Orchestration: Never experience a "Debugger already attached" race condition again. The extension enforces deterministic lock ownership per tab. If Agent A locks Tab 1, Agent B is actively blocked and receives a structured lock conflict error, allowing it to queue or wait.

Rich Diagnostic UI: A neo-industrial popup interface gives you real-time telemetry. View active connections, inspect which agent owns which tab lock, copy your bridge URL, and forcibly terminate runaway agent sessions directly from the browser.

Resilient Crash Recovery: Built for the realities of AI development. If an agent crashes while holding a tab lock, or if the extension restarts, the system auto-reconciles stale state, clears ghost locks, and reconnects without requiring manual cleanup.

🛠️ Exposed MCP Capabilities

When connected to the Local MCP Server, this extension exposes a massive surface area of browser automation tools directly to your LLM:

Tab Management: list_available_tabs, attach_to_tab, detach_from_tab (lock-aware orchestration).

Navigation & State: browser_navigate (URL, back, forward, reload), browser_tabs (new, close, attach).

DOM Intelligence: browser_snapshot (structured DOM tree extraction), browser_extract_content (targeted text extraction), browser_lookup.

Execution & Interaction: browser_evaluate (execute raw JS via Runtime.evaluate), browser_interact (click, type, hover, scroll), browser_fill_form, browser_drag.

Network & Diagnostics: browser_network_requests (capture and filter XHR/Fetch via event listeners), browser_console_messages, browser_performance_metrics.

Media: browser_take_screenshot, browser_pdf_save (export pages to base64 PDF).

Environment Control: browser_window (resize, maximize), browser_handle_dialog, browser_list_extensions, browser_verify_text_visible.

🎯 Who is this for?

Software Engineers using IDEs like Cursor who want their AI to test web applications, reproduce bugs, or read local documentation.

AI Operators building autonomous web-scraping or research agents using frameworks like LangChain, AutoGen, or raw MCP clients.

QA Automation Engineers who want to leverage LLMs to write and execute browser interactions dynamically without writing fragile XPath selectors.

⚠️ Important Setup Requirements

This extension does not operate on its own. It is the browser-side bridge for the Local MCP Bun Server.

To use this extension, you must run the local server on your machine:

Download the server binary from the project repository (https://github.com/TWP-Technologies/Browser-MCP/releases).

Run the server locally (defaults to ws://127.0.0.1:37777/extension).

Configure your MCP client (e.g., claude_desktop_config.json or Cursor's MCP settings) to point to the local server.

The extension will automatically connect, pulse to indicate readiness, and begin routing agent commands.

(Modified by [KnotFalse] - Originally inspired by the blueprint-mcp architecture, completely rebuilt in Bun/TypeScript for local-first, high-performance multiplexing).