Agent Desktop by Lahfir is a free MCP server tool built with Rust, designed for desktop automation by AI agents. It provides structured access to native applications via OS accessibility trees and connects to Chromium-based applications using the Chrome DevTools Protocol. Distributed as prebuilt C-ABI cdylibs and CLI tarballs, it supports macOS, Linux, and Windows, enabling AI clients like Claude Desktop and Cursor to interact with local desktop environments. Developers building automation tools or research prototypes will find it useful.

What does Agent Desktop connect to?

Agent Desktop connects AI agents to both native desktop applications and Chromium-based applications. For native applications, it uses OS accessibility trees to provide structured access. It connects to Chromium-based applications like Slack, VS Code, Discord, Obsidian, and Notion by exposing a Chrome DevTools Protocol endpoint. As an MCP server tool, it supports local process communication via stdio and remote communication through Streamable HTTP, replacing the deprecated HTTP+SSE transport.

Core features and tools

Agent Desktop exposes 53 commands for inspecting and operating native applications, all with JSON output. It also includes specific tools for session tracing: trace show and trace export. These capabilities allow AI agents to control any native desktop application and drive Chromium-based applications effectively. Compatible AI clients include Claude Desktop, Claude Code, ChatGPT (Web), and Cursor App, with frameworks like LangChain and LlamaIndex also supported.

Who is it for?

Agent Desktop targets AI agents, developers creating desktop agents, internal automation tools, and research prototypes. It enables AI agents to control native applications through OS accessibility trees, providing structured JSON output. It also allows AI agents to drive Chromium-based applications via the Chrome DevTools Protocol. This tool offers a standardized MCP server interface for various AI clients to interact with the local desktop environment. Skip this if your project does not involve AI agent interaction with desktop applications.

How to get started

MCP servers generally require authentication, supporting key-based methods, Microsoft Entra, and OAuth identity passthrough. Setup typically involves configuring credentials and adding the server URL to the client's MCP settings, with eager authentication occurring once per session.

Use cases

• Enabling AI agents to control any native desktop application through OS accessibility trees with structured JSON output.
• Allowing AI agents to drive Chromium-based applications by connecting via the Chrome DevTools Protocol.
• Providing a standardized MCP server interface for AI clients (e.g., Claude Desktop, Cursor) to access and interact with the local desktop environment.

Agent Desktop offers a powerful interface for AI agents to interact with desktop environments, both native and Chromium-based. Its Rust foundation and MCP server design make it a capable tool for automation and agent development. It provides structured access and command execution with JSON output, supporting a range of AI clients and frameworks. Start by exploring its capabilities for controlling a Chromium-based application like Slack.