Speak Blueprint.

Aethyr is an MCP server that lets your AI assistant read and author Unreal Engine 5 Blueprints in plain language. This page gets you from zero to your first AI-made edit in about five minutes.

New here UE veteran Both of you are covered: beginner steps are spelled out; the deeper "how it works" notes are marked for the vets.

What it actually is

Two pieces that travel together:

A UE editor plugin that knows how to introspect and safely mutate Blueprint assets.
A standalone MCP server (AethyrMcp) that speaks the Model Context Protocol, so AI clients can drive that plugin by talking to it: no Python, no editor scripting, no C++.

This build ships 82 tools, deliberately the subset we're confident works without any chance of corrupting your project. Reads can't mutate anything; every write saves through a compile check and an automatic backup.

5-minute quickstart

Get the plugin. Install Aethyr from Fab Fab into your project's Plugins/ folder and enable it. New Restart the editor when prompted.
Point it at your project. The server needs to know which .uproject to drive. It auto-discovers one above its own folder, or you set AETHYR_PROJECT (next step).
Connect your AI client over MCP (below).
Ask it something. Try: "List the Blueprints under /Game and tell me what BP_PlayerCharacter does."

That's the whole loop. From here, everything is just asking for what you want. See the user-story tutorials.

Connect your AI

Aethyr is a standard MCP stdio server. Add it to your client's MCP config. The shape is the same everywhere, only the file location differs.

Claude Code / Claude Desktop / Cursor / Copilot

// mcpServers block (claude_desktop_config.json, .cursor/mcp.json, .mcp.json, etc.)
{
  "mcpServers": {
    "aethyr": {
      "command": "C:/Path/To/AethyrMcp.exe",
      "env": {
        "AETHYR_PROJECT": "C:/Path/To/YourGame.uproject",
        "AETHYR_ALLOW_WRITE": "1"
      }
    }
  }
}

Read-only by default. The server refuses every write tool until you set AETHYR_ALLOW_WRITE=1, so an exploratory session can't change anything by accident. Leave it off until you want edits.

Vet The backend is auto by default: it probes per call and routes to a live editor if one's open (over a local socket), or spins a headless editor when it isn't. You never pick a backend by hand.

Skills & the audit agent

The plugin ships Claude skills and an agent that teach your AI how to use the tools well. Install them with the bundled script (Install-ClaudeAssets.ps1), or point your client's skills folder at the plugin's Claude/ directory.

aethyr: the core skill: reading and editing a Blueprint, the wire format, per-tool guidance.
bp-batches: multi-step edits as one atomic transaction (apply_ops / preview_ops / compile_function).
bp-debug: what to do when a call fails: error codes, recovery, partial-write cleanup.
Agent bp-audit: a read-only structural auditor. Ask it "audit /Game/AI" and it fans out reads and returns a tight report without touching anything.

Why this matters: the skills are what turn "call a tool" into "do the right sequence of tools." With them installed, your AI reaches for preview_ops before a risky batch, keys links on (node, pin), and explains a failure instead of flailing.

How it works (for the curious)

Vet A few load-bearing ideas:

Progressive vs. full surface. This 82-tool build advertises its whole surface on tools/list, small enough that your AI sees everything at once.
Corruption-safety is structural. Every kept write flows through CompileAndSaveBlueprint: it won't save a graph that failed to compile (unless you opt in), and it copies the asset to a backup ring before the first save of a session.
Atomic batches. apply_ops with atomic:true reverts every change in the batch if any step fails, so your asset is never left half-rewritten.
Package paths, not object paths. Everything addresses assets as /Game/AI/BP_Foo, never /Game/AI/BP_Foo.BP_Foo.