This project is a workbench that lets an AI agent (Claude Code) and its operator drive a live Unreal Editor. Its trust boundary is your local machine, and it is designed to run with your own privileges.

Two listeners make up the harness, and both bind loopback by default:

Neither listener is authenticated. By design, the bridge executes operator-authored code — typed tool calls, and the deliberate editor_console_exec escape hatch (including its py path) run arbitrary console/Python inside your editor. Treat anything that can reach these ports as able to act as you, with your permissions, on your machine and project.

Remote code execution is therefore not a vulnerability in this model — running the operator's code is the feature. The security posture is entirely about who can reach the ports.

• Do not bind the listeners to a routable interface. UNREAL_MCP_HOST and UNREAL_BRIDGE_HOST can be overridden (e.g. to 0.0.0.0); doing so exposes an unauthenticated, code-executing surface to your network. Leave them on 127.0.0.1.
• If you must reach the server from another host, put it behind an authenticated tunnel (SSH, a reverse proxy with auth, a VPN) rather than binding off-loopback.
• The harness is intended for a single trusted operator on the local machine, not as a multi-tenant or internet-facing service.

If you find a security issue that falls outside this local-trust model — for example a default that unexpectedly binds off-loopback, a path that escapes the intended sandbox, or a dependency advisory — please report it privately:

• Open a GitHub Security Advisory on this repository ("Security" tab → "Report a vulnerability"), or
• Open a regular issue for non-sensitive hardening suggestions.

Please do not open a public issue for something exploitable until it has been addressed.

If you aren't creating assets and et cetera yourself, you are probably downloading untrustable binary data formats for a source-available engine that basically does RCE by design. Luckily, it's all fun and games. I wouldn't host the UE editor in prod, though. The games should be hosted in environments that are sacrificial and isolated.

PRECISELY NOTHING ABOUT THE UNREAL EDITOR IS SECURE. I run this most of the time with --dangerously-skip-permission. Buyer beware, caveat emptor, no liability for damages you incur on yourself. Don't gamble with LLMs you can't afford doing something wacky.