Auten

What the Auten MCP server does

Auten MCP server gives an MCP-compatible agent control over applications running on a local computer. The agent can inspect visible interfaces, identify controls, open applications or URLs, enter text, select menus, scroll, drag, and press keys. Supported desktop systems are macOS, Windows, and Linux. Android control is available in beta through a USB-connected phone.

The server is intended for software that does not provide an API. Example workflows include exporting a report from a desktop application, downloading files from a supplier portal, moving information between a CRM and email, or replying to a message in a phone application.

Auten MCP server also supports reusable skills. After an agent completes a workflow, it can save that task and later replay it with new values. Replay is performed without another model call, and the material states that the replay can adapt when an application's layout changes. Stored credentials can be entered with fill_login, so the secret is typed locally rather than exposed to the agent.

How it works

A small runner operates on the computer being controlled. The MCP client communicates with that runner through the auten mcp-bridge command or the npm launcher npx -y @autenai/mcp. The runner reads the accessibility tree when possible, providing named or numbered interface elements for actions. Screenshots, OCR, and vision-related tools cover interfaces with limited accessibility data.

The normal connection uses MCP over stdio. A runner-only mode is also available for network clients using SSE; in that arrangement, the runner remains online on the target computer and the client uses a personal MCP URL. The README warns that this URL contains a key and should remain private.

Setup and configuration

Install the runner with the macOS/Linux shell command or the Windows PowerShell command documented by the project. The first run asks you to connect a free key. Then register auten mcp-bridge with the chosen MCP client, or use the npx package entry when the client can launch npm packages.

The README provides setup instructions for Claude Code, Codex, Cursor, Gemini CLI, Claude Desktop, VS Code agent mode, Windsurf, Zed, and Cline. Cursor and VS Code can also use the provided install links. On Windows, clients that cannot launch npx directly may need cmd /c npx -y @autenai/mcp, or the direct runner command.

Tools and capabilities

The documented tool groups include:

• Screen inspection: list_elements, read_text, screenshot, screenshot_labeled, vision, and screen_info.
• Input and interaction: click_at, click, fill, type, paste_text, press_keys, scroll, drag, move_mouse, act_on_matching, and menu_select.
• Application control: open_app, open_url, and focus_window.
• Credential entry through fill_login.
• Saving and replaying completed workflows as skills.

Limitations and notes

Android support is marked beta and requires USB connectivity. The runner must be installed or available on the computer whose applications it will control. Auten does not provide its own AI model; the connected agent supplies the planning and decision-making. An agent using the hosted MCP connection does not need to provide its own AI key, while the standalone CLI can be given a key for supported providers with auten key add.

The server acts on visible applications rather than calling their APIs, so the agent may need to inspect the interface and choose actions interactively. Accessibility support varies by application; screenshots and OCR are available for applications without sufficient accessibility information.