Astrax, a new open-source Model Context Protocol (MCP) server written in Python, has launched to give AI agents direct control over Windows desktops. Developed by Rufus (@AstraxOf), the tool allows Large Language Models (LLMs) to see, click, type, and execute multi-step tasks without saturating the model's context window. It is compatible with major coding assistants and AI clients, including Claude Desktop, Roo Code, Cline, Cursor, Windsurf, and Gemini CLI.

Solving the Context Saturation Problem

Astrax addresses a critical pain point for developers using local models via Ollama: context saturation. When AI agents perform complex, multi-step desktop automation, the accumulation of screenshot history and previous actions often leads to hallucinations. Astrax mitigates this by isolating task context in a local state file, .astrax_task_state.json. Instead of re-reading the entire conversation history, the AI retrieves only the current step, the next step, and essential data, such as file paths or IDs, ensuring lightweight and accurate execution.

Granular Permissions and Security

Security is handled through a single-authorization model per task. Before execution, the AI declares required permissionsβ€”such as pantalla (screen), raton_teclado (mouse/keyboard), portapapeles (clipboard), or terminalβ€”and requests user confirmation once. Once authorized, the agent operates for a default 60-minute session. The tool includes fail-safes, such as a mouse-abort mechanism where moving the cursor to a screen corner stops automation, and blocking of destructive shell commands like rm -rf / or format C: to prevent accidental data loss.

One-Click Installation and Compatibility

Installation is streamlined via uv, the fast Python package manager. A single command detects the operating system and registered MCP clients, automatically injecting the Astrax configuration. The installer supports dry runs, specific client targeting (e.g., --client claude-desktop), and auto-approval flags for clients like Roo Code. While primarily marketed for Windows, Astrax supports macOS (requiring screen recording permissions) and Linux (with X11), though Wayland support is currently limited due to capture restrictions.

Key Takeaways

  • Astrax isolates AI context in local JSON files to prevent hallucinations in long-running desktop tasks.
  • Supports one-click installation via uv for Claude Desktop, Cursor, Cline, and other MCP clients.
  • Implements a single-authorization security model with granular permissions for mouse, keyboard, and terminal access.
  • Compatible with local models (Ollama) and commercial APIs (Claude, Gemini) via stdio transport.

The Bottom Line

Astrax is a pragmatic solution for the 'context rot' that plagues agentic desktop automation. By treating state as a first-class citizen in local files rather than relying on fragile chat history, it offers a robust path for developers wanting to deploy reliable, long-running AI workflows on Windows.