Unreal MCP Server 3.0: 359 Tools and a Closed-Loop AI Agent for UE5

Today we're shipping Unreal MCP Server v3.0.0-alpha.12 — a release that turns the plugin from "AI co-pilot for editor scripting" into a true closed-loop editor agent platform.

The headline numbers: 359 tools (up from 305), 15 resources, 11 prompts, organized across 50+ categories. But the tool count understates what changed. v3 adds three brand-new capability surfaces — PIE control, source control, and runtime debug — that together let an AI agent write code, hit Play, observe what happened, and fix itself, without a human in the loop until something needs review.

Launch pricing. Personal stays at $69.99 and Professional stays at $199.99 for the next 5 days. On 2026-05-07 they rise to $89.99 and $299.99 respectively. Existing v2 customers get v3 free — your license covers the entire 3.x line. → Lock in the launch price

The closed loop

This is the demo we kept showing internally and that finally convinced us to ship:

agent: tools/call create_blueprint_event
agent: tools/call pie_start { mode: "Selected" }
agent: tools/call pie_send_input { key: "SpaceBar", event: "Pressed" }
agent: tools/call pie_screenshot
agent: tools/call get_last_runtime_error
   ⤷ "AccessNone in BP_Player.OnJump → SkeletalMesh"
agent: tools/call edit_blueprint_function
agent: tools/call pie_stop
agent: tools/call pie_start { mode: "Selected" }
agent: ✅ no error this time

The agent runs that loop on its own. The nine new pie_* tools (pie_start, pie_stop, pie_pause, pie_resume, pie_step_frame, pie_send_input, pie_screenshot, pie_get_state, pie_attach_player_controller) are backed by GEditor->RequestPlaySession and APlayerController::InputKey via FInputKeyEventArgs::CreateSimulated. See the PIE Control docs for the full tour.

Pair PIE with the new Test Authoring tools (create_automation_spec, run_automation_specs, get_last_test_report, …) and the agent can scaffold and run UE automation specs end-to-end.

Source-control aware agents

Studios reject AI editor agents for one reason above all others: traceability. The eight new sc_* tools close that gap. They wrap ISourceControlModule, so they work against whatever provider the project has loaded — Perforce, Git LFS, Plastic, or Subversion — without the agent needing to know which.

{ "method": "tools/call", "params": {
  "name": "sc_check_out",
  "arguments": { "files": ["/Game/Blueprints/BP_Door.uasset"] }
} }

// ... agent edits ...

{ "method": "tools/call", "params": {
  "name": "sc_submit",
  "arguments": {
    "files": ["/Game/Blueprints/BP_Door.uasset"],
    "description": "agent run #4711: wire OnInteract → PlayDoorOpenSound"
  }
} }

That's the difference between "someone changed BP_Door" and "agent run #4711 modified BP_Door at 14:32 with description X". Read the full Source Control walkthrough.

Production-safe by default

v3 (Phase B) introduces the security model studios kept asking for:

• Bearer-token auth. Set bRequireAuthToken=true and clients must send Authorization: Bearer <token>. Missing or invalid → 401.
• Origin allow-list. Default [http://localhost, http://127.0.0.1]. Non-allowed origins get 403. Replaces v2's Access-Control-Allow-Origin: * — the one breaking change in v3.
• Read / Scene / Destructive scopes. Every tool refuses below its required scope. Set bAllowDestructiveScope=False in CI environments and no caller can ever delete or submit, even with a valid token.

Plus three agent-ergonomics primitives that make autonomous loops actually work:

• Multi-call transactions (transactions/begin|commit|rollback) collapse a sequence of edits into one Ctrl+Z step.
• Dry-run wraps an opt-in tool call in a cancelled FScopedTransaction so the agent can preview an effect.
• Structured EMCPError taxonomy with did-you-mean fuzzy suggestions on not_found. The agent branches on code, not text-matching message.

Full reference at Security and Agent Ergonomics.

What's new at a glance

Plus Phase A (new MCP_TOOL builder + FMCPValidate validator layer; ~3000 LOC eliminated from tool registrations) and the architectural Phase B + C work above.

Migration from v2

For 99% of users: drop in the new plugin, restart the editor, you're done. Two things to watch:

1. CORS. If you connect from a non-localhost browser, add the origin to AllowedOrigins.
2. Engine version. v3 is Unreal Engine 5.7 only. UE 5.6 users should stay on v2.0.2 until you upgrade.

Existing v2 license keys unlock v3 automatically — no re-purchase, no upgrade fee.

What's deferred to v3.1

A handful of D.x bodies ship as schema-only Unsupported in alpha.12 — registered surface, no implementation yet. We chose this deliberately to keep the plugin loadable on projects without specific deps:

• Modeling tool bodies (need MeshModelingToolsExp + ModelingComponents + GeometryCore)
• MetaSound graph mutators (need MetasoundFrontend)
• Chaos asset creation tools (need GeometryCollectionEngine + FractureEngine + ChaosCloth)
• metahuman_import (needs Quixel Bridge Python entry points)
• WebSocket transport (needs vendored Boost.Beast or libwebsockets)
• Live notifications/progress streaming (needs async tool execution)

These are tracked on the Q3 2026 roadmap and will land in the v3.1 cycle.

Get it

→ Lock in launch pricing — 5 days only. → Read the v3 docs → Full changelog

If you want to see the closed loop in action, follow the PIE control tutorial.