glass — Claude Code MCP server: what it does, how to install
Connects your coding agent to native desktop and mobile GUI applications. It runs the software, drives input, and checks UI states so the model can fix bugs autonomously.
Open on GitHubWhat glass does
glass treats native software as an external black box. It skips routine screenshots entirely. The server runs binaries, captures frames, injects input, and collects application logs. When inspecting widgets, the agent queries the accessibility tree using glass_a11y_snapshot. For raw graphics, glass_diff computes pixel delta percentages and outputs bounding boxes in text.
The core is written in Rust. It supports X11, Wayland, Windows, macOS, alongside Android emulators and the iOS Simulator. Host input isolation depends on the target environment. On Windows and macOS, simulated clicks share the physical desktop, while Linux virtual displays and mobile emulators stay isolated. If prompt tokens are tight, glass-mcp --tool-profile lean merges multiple tools into glass_do.
What it helps with
- debug native desktop bugs without manual screenshot approvals
- inspect UI widgets and control states through the accessibility tree
- measure visual updates via pixel diffs to reduce vision token spend
- drive automated test flows inside Android emulators and the iOS Simulator
Example prompts for the agent
- Run
tasks_demo.pyunder glass and trace why clicking Add does not insert a new task. - Take an accessibility snapshot of the current window and fill out the username field.
- Click the Save button using
glass_doand verify the success notification in logs.
How to install glass
Install
The MCP registry has no launch command for it: see the repository README
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about glass
How do I install glass in Claude Code?
Download the prebuilt binary from the repository releases or compile from source using cargo build. Link the binary inside your Claude Code MCP settings, then run glass-mcp doctor to verify system requirements.
What does glass do?
It provides an interaction loop for coding agents working on graphical apps. The agent launches the binary, simulates clicks or keystrokes, reads logs, and confirms visual state changes on its own.
Does glass capture my personal desktop cursor?
On native macOS and Windows builds, mouse and keyboard actions share your active desktop session. Virtual Linux displays and mobile emulators isolate input completely.
