MCP Browser Agent

What imprvhub/mcp-browser-agent MCP server does

The imprvhub/mcp-browser-agent MCP server exposes browser control and HTTP request operations to an MCP client, with Claude Desktop as the documented client. Browser tools can navigate to URLs, capture screenshots, click elements, fill form fields, choose dropdown options, hover over elements, and execute JavaScript in the browser context. The JavaScript capability also captures console logs.

Separate API tools send GET, POST, PUT, PATCH, and DELETE requests. Requests can include headers and body content, and responses are processed with JSON formatting and error feedback. This combination supports workflows that move between a rendered website and an endpoint exposed by the same or another service.

How it works

The project uses Playwright for browser automation and maintains a persistent, headful browser session. An MCP client sends tool calls to the local process, which performs the requested browser or HTTP operation and returns the result. Screenshots and browser console logs can also be accessed through MCP resources.

The browser can be selected as Chrome, Firefox, WebKit, or Edge. Chrome is the default. WebKit is the rendering engine used in place of direct Safari support. Playwright may install the relevant browser driver when that browser type is first used, or the drivers can be installed manually.

The imprvhub/mcp-browser-agent MCP server is suited to chained tasks such as opening a page, completing a form, submitting it, and capturing the resulting page. The README also describes natural-language task execution and error recovery for multi-step operations, but the exact recovery behavior depends on the task and page being automated.

Setup and configuration

The documented prerequisites are Node.js 16 or newer, Claude Desktop, and Playwright dependencies. Installation requires cloning the repository, running npm install, and building with npm run build. The server can then be started from the project directory with node dist/index.js.

Claude Desktop can start the process through its claude_desktop_config.json file. The configuration uses node as the command and points its arguments to the built dist/index.js file, with an optional browser argument such as --browser chrome. The configuration path differs by macOS, Windows, and Linux, so the repository directory must be supplied as an absolute path.

Browser selection is also configurable through .mcp_browser_agent_config.json in the home directory or through the MCP_BROWSER_TYPE environment variable. Supported values are chrome, firefox, webkit, and edge. A command-line --browser option is another supported method.

Tools and capabilities

The imprvhub/mcp-browser-agent MCP server provides these browser tools:

• browser_navigate for opening a URL with configurable load behavior.
• browser_screenshot for full-page or element-specific captures.
• browser_click, browser_fill, browser_select, and browser_hover for DOM interactions.
• browser_evaluate for running JavaScript in the page and collecting console logs.

Its HTTP tools are api_get, api_post, api_put, api_patch, and api_delete. These support request headers and body content and return formatted response data with error details. MCP resources expose console logs and screenshots.

Limitations and notes

Safari is not directly supported; selecting webkit uses Playwright’s WebKit engine instead. The Edge option launches Microsoft Edge through Playwright’s Edge channel rather than providing a separate browser engine. Browser drivers and the target browser must be available or installed before the corresponding automation can run.

The documented client is Claude Desktop, and the server runs locally rather than through a hosted endpoint. The provided material does not specify an authentication mechanism, package registry release, or software license.