Agent Web Interface

An MCP server for browser automation that exposes semantic, token-efficient page representations optimized for LLM agents.

Motivation

LLM-based agents operate under strict context window and token constraints. However, most browser automation tools expose entire DOMs or full accessibility trees to the model.

This leads to:

• Rapid token exhaustion
• Higher inference costs
• Reduced reliability as relevant signal is buried in noise

In practice, agents spend more effort finding the right information than reasoning about it.

Agent Web Interface exists to change the unit of information exposed to the model.

Core Idea: Semantic Page Snapshots

Instead of exposing raw DOM structures or full accessibility trees, Agent Web Interface produces semantic page snapshots.

These snapshots are:

• Compact and structured
• Focused on user-visible intent
• Designed for LLM recall and reasoning, not DOM completeness
• Stable across layout shifts and DOM churn

The goal is not to mirror the browser, but to present the page in a form that aligns with how language models reason about interfaces.

How It Works

At a high level:

1. The browser is controlled via Puppeteer and CDP
2. The page is reduced into semantic regions and actionable elements
3. A structured snapshot is generated and sent to the LLM
4. Actions are resolved against stable semantic identifiers rather than fragile selectors

This separation keeps:

• Browser lifecycle management isolated
• Snapshots deterministic and low-entropy
• Agent reasoning predictable and efficient

Benchmarks

Early benchmarks against Playwright MCP show:

• ~19% fewer tokens consumed
• ~33% faster task completion
• Same or better success rates on common navigation tasks

Benchmarks were run using Claude Code on representative real-world tasks. Results are task-dependent and should be treated as directional rather than absolute.

What Agent Web Interface Is (and Is Not)

Agent Web Interface is:

• A semantic interface between browsers and LLM agents
• An MCP server focused on reliability and efficiency
• Designed for agent workflows, not test automation

Agent Web Interface is not:

• A general-purpose browser
• A visual testing or screenshot framework
• A replacement for Puppeteer

Puppeteer remains the execution layer; Agent Web Interface focuses on representation and reasoning.

Usage

Agent Web Interface implements the Model Context Protocol (MCP) and works with:

• Claude Code
• Claude Desktop
• Cursor
• VS Code
• Any MCP-compatible client

Example workflows include:

• Navigating complex web apps
• Handling login and consent flows
• Performing multi-step UI interactions with lower token usage

Claude Code

# Default: connects to your running Chrome, falls back to launching a new browser
claude mcp add agent-web-interface -- npx agent-web-interface@latest

# Explicit mode: always connect to your Chrome (no fallback)
claude mcp add agent-web-interface -e AWI_BROWSER_MODE=user -- npx agent-web-interface@latest

# Explicit mode: always launch a persistent browser
claude mcp add agent-web-interface -e AWI_BROWSER_MODE=persistent -- npx agent-web-interface@latest

CLI Arguments

The server accepts transport-level arguments only. Browser configuration is per-session via the navigate tool.

| Argument | Description | Default | | ------------- | --------------------------------- | ------- | | --transport | Transport mode: stdio or http | stdio | | --port | Port for HTTP transport | 3000 |

Browser Session Modes

Browser initialization is automatic on the first tool call. Set AWI_BROWSER_MODE to control how the browser is started:

| Mode | Behavior | Profile | | ------------ | ---------------------------------------------------- | --------------------------------------------- | | (unset) | Auto: try user → persistent → isolated | Depends on fallback | | user | Connect to your running Chrome | Chrome's default profile | | persistent | Launch Chrome with a dedicated persistent profile | ~/.cache/agent-web-interface/chrome-profile | | isolated | Launch Chrome with a temporary profile (clean slate) | None (deleted on close) |

Examples:

# Auto mode (default) — connects to Chrome if running, else launches
npx agent-web-interface

# Always connect to your Chrome (fails if Chrome isn't running)
AWI_BROWSER_MODE=user npx agent-web-interface

# Always launch with persistent profile
AWI_BROWSER_MODE=persistent npx agent-web-interface

# Headless isolated browser
AWI_BROWSER_MODE=isolated AWI_HEADLESS=true npx agent-web-interface

# HTTP transport mode
npx agent-web-interface --transport http --port 8080

Using Your Existing Chrome (Chrome 144+)

To connect with your bookmarks, extensions, and logged-in sessions:

1. Navigate to chrome://inspect/#remote-debugging in Chrome
2. Enable remote debugging and allow the connection
3. Set AWI_BROWSER_MODE=user (or leave unset for auto mode)

{
  "mcpServers": {
    "agent-web-interface": {
      "command": "npx",
      "args": ["agent-web-interface@latest"],
      "env": {
        "AWI_BROWSER_MODE": "user"
      }
    }
  }
}

Environment Variables

| Variable | Description | Default | | ------------------ | -------------------------------------------------------- | ------------------------ | | AWI_BROWSER_MODE | Browser mode: user, persistent, or isolated | (unset: auto fallback) | | AWI_HEADLESS | Run browser headless (true/false) | false | | AWI_CDP_URL | Explicit CDP endpoint (overrides mode) | - | | AWI_TRIM_REGIONS | Set to false to disable region trimming globally | true | | TRANSPORT | Transport mode override (http) | - | | HTTP_HOST | Host for HTTP transport | 127.0.0.1 | | HTTP_PORT | Port for HTTP transport | 3000 | | LOG_LEVEL | Logging level | info | | CEF_BRIDGE_HOST | CDP host for CEF bridge connection | 127.0.0.1 | | CEF_BRIDGE_PORT | CDP port for CEF bridge connection | 9223 | | BRING_TO_FRONT | Set to true to focus the Chrome tab before each action | false | | CHROME_PATH | Path to Chrome executable (multi-tenant) | - |

Installation

git clone https://github.com/lespaceman/agent-web-interface
cd agent-web-interface
npm install
npm run build

Configure the MCP server in your client according to its MCP integration instructions.

Architecture Overview

Agent Web Interface separates concerns into three layers:

• Browser lifecycle — page creation, navigation, teardown
• Semantic snapshot generation — regions, elements, identifiers
• Action resolution — mapping agent intent to browser actions

This separation allows each layer to evolve independently while keeping agent-visible behavior stable.

Status

Agent Web Interface is under active development. APIs and snapshot formats may evolve as real-world agent usage informs the design.

Feedback from practitioners building agent systems is especially welcome.

License

MIT