# browser-agent [Health: Active]

**Category:** 📂 Browser Automation  
**Repository:** https://github.com/runbook-ai/runbook-ai-mcp  
**GitHub Stars:** 3  
**npm Downloads (last month):** 213  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/browser-agent

## Description
MCP server for browser automation with AI agent

## Tools
Capabilities this server exposes over MCP:

- **browser-agent** — Run a task in Chrome browser with AI and automation capabilities.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "browser-agent": {
    "command": "npx",
    "args": ["-y","runbook-ai-mcp@latest"],
    "env": {
      "RUNBOOK_AI_MCP_TOKEN": ""
    }
  }
}
```

**Requires environment variables:** `RUNBOOK_AI_MCP_TOKEN` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What browser-agent MCP server does

The browser-agent MCP server lets an MCP client ask an AI agent to complete tasks in a live Chrome browser. It uses a Chrome extension to perform actions such as navigation, clicking, typing, page inspection, and other browser automation steps. The extension can work with websites that do not provide a dedicated MCP integration, including public sites, account pages, and internal web tools.

The server exposes one MCP tool, `browser-agent`. Its required `prompt` describes the task in ordinary language. The tool can return task results while placing generated files, downloads, extracted datasets, screenshots, or captured payloads on the machine running the MCP server. File paths, MIME types, and sizes are listed in the result instead of placing file contents directly in the MCP response.

## How it works

The MCP client communicates with the local server over stdio. When the tool is called, the server forwards the request over a WebSocket connection on port 9003 to the Runbook AI Chrome extension. The extension runs the browser task and sends the result back through the server.

Chrome must be open with the extension side panel available. The extension generates a reduced representation of page HTML that keeps relevant text and interaction elements while omitting much of the surrounding markup. This representation can include content beyond the current viewport, which supports scrolling and multi-page tasks without returning the complete DOM after every action.

Each call is isolated by default: it starts in a fresh blank tab and closes tabs opened during the run. Set `ephemeral` to `false` when later calls should continue from the previous browser state and leave the final page open. `maxIterations` controls the maximum number of agent actions, while `effort` selects quick, normal, or thorough exploration.

## Setup and configuration

Install the MCP package through an MCP client configuration using `npx -y runbook-ai-mcp@latest`. Install the Runbook AI extension from the Chrome Web Store, open its settings, and enable MCP. The extension also requires an LLM API key, model name, and base URL; these values are configured in the extension rather than as MCP server environment variables.

The extension accepts connections on localhost:9003. Without a token, any program able to open a WebSocket to that port can drive the browser. Generate an MCP token in the extension settings and pass it with the server's `--token` argument to restrict requests. The server also accepts `RUNBOOK_AI_MCP_TOKEN` when the argument is not supplied; this avoids exposing the token in process listings on shared machines. An empty token disables the check.

By default, the extension can access all websites. Chrome's extension details can be used to restrict access to selected sites. The browser-agent MCP server does not itself make remote calls except to the LLM provider selected in the extension configuration.

## Tools and capabilities

- Run one natural-language browser task with the `browser-agent` tool.
- Set an iteration limit with `maxIterations`; longer workflows require more iterations and token budget.
- Choose `quick`, `normal`, or `thorough` exploration with `effort`.
- Preserve or isolate browser state with `ephemeral`.
- Write files to a chosen `outputDir`, or use the default temporary location.
- Return paths for saved files and show images up to 4 MB inline as MCP image content.

The server fits the Browser Automation category when an agent needs to operate a website through Chrome rather than call a site-specific API.

_Full upstream README: https://allmcps.com/mcp/browser-agent/readme_