Skip to content
github-actions[bot] edited this page Oct 19, 2026 · 1 revision

Safety

한국어

  • Local only. The server listens on 127.0.0.1. Requests whose browser Origin is not localhost, 127.0.0.1 or [::1] are refused. AuthToken additionally requires a bearer token.
  • Undoable, all-or-nothing writes. Write tools validate values before changing anything and run in an editor transaction. If a tool fails after changing state, the transaction is rolled back.
  • No writes during play. Write tools and asset creation/import tools are refused while Play In Editor is starting or running.
  • Project content only. Assets can be created, imported, changed and saved only under /Game and project plugins. Engine content is read-only.
  • Explicit saving. No tool saves as a side effect, with one exception: level_new writes the level it creates, because a level only exists on disk. asset_save saves loaded project assets only and still refuses levels; levels are saved with level_save, which writes the level open in the editor.
  • No level is replaced silently. level_new and level_open refuse to run while a level has unsaved changes, because opening a level discards them.
  • Dry runs. Destructive tools need bConfirm: true to act, and asset_import_textures validates every entry before importing anything.
  • Allow and block lists. AllowedTools and BlockedTools hide tools, while BlockedProperties protects properties from object_set_properties and UMG tools.
  • Files. skills_get can read only inside a skill folder and refuses outside paths and hidden files.
  • Source files. The source tools read and write text files only inside SourceFolders and refuse other paths, .. that climbs out, build folders and files that are no text. Their edits are not undoable from the editor.
  • Approval for tagged clients. A session that sends the X-AgentMcp-Client header keeps that tag for its lifetime. Every tool call of a tagged session other than a Read tool waits for the approval gate registered for the tag; with no gate registered it fails with APPROVAL_UNAVAILABLE, and a refusal returns DENIED. The AI chat registers the gate agent-mcp-chat and shows an approval card for each change. Sessions without the header are not affected, so Claude Code and Codex connected directly work as before.
  • Perf files. The perf tools read files only inside the project folder. perf_trace_hotspots starts the engine's own UnrealInsights.exe to export a trace next to it.
  • There is no tool that runs console commands or arbitrary scripts.

Clone this wiki locally