egui_mcp

An MCP server that lets an AI agent (Claude, Codex, …) drive a live egui app.

egui-mcp connects to a running app over the egui_inspection protocol and exposes it as MCP tools: read the AccessKit widget tree (widget_tree / get_widget), click / type / scroll / drag / press keys, take a screenshot, resize the window, and wait_for async UI to settle.

1. Enable inspection in your app

Enable eframe's inspection feature and launch the app with the EGUI_INSPECTION env var set:

EGUI_INSPECTION=1 cargo run --features eframe/inspection   # binds 127.0.0.1:5719

When the variable is unset or falsy (0 / false), inspection is completely off. See egui_inspection for details, including how to expose it across the network.

2. Install the mcp

cargo install --git https://github.com/rerun-io/egui_crates egui_mcp

This installs the egui-mcp binary onto your PATH.

3. Configure your agent

Claude Code — register the server with one command:

claude mcp add egui egui-mcp

or add it to your MCP config (~/.claude.json or .mcp.json) manually:

{
  "mcpServers": {
    "egui": {
      "command": "egui-mcp"
    }
  }
}

Codex — add it to ~/.codex/config.toml:

[mcp_servers.egui]
command = "egui-mcp"
args = []

Using it

With the app running and the agent configured, ask the agent to attach, then drive the app. You could ask it to e.g. reproduce a bug, verify a new feature or just randomly use, test and explore your app.

Screenshots need a visible window. Reading the tree and injecting input work even while the app is in the background, but capturing a screenshot requires a rendered frame — which the OS won't produce for a fully-occluded or minimized window (notably on macOS). Bring the window to the foreground to capture it.

Developing egui_mcp