UEREMCP

A goal-level agent interface for Unreal Engine 5.8.

An AI describes a complete outcome — "create a replicated fireball ability with VFX, materials, gameplay tags, damage, and animation" — submits one structured JSON request, and receives one validated structured response with a complete change manifest. Instead of several hundred primitive editor calls.

Status: ready for practical goal-level use within cataloged scopes (2026-07-30), not production-perfect. Create/inspect surfaces and deterministic Niagara frame capture are live-verified against RE, including fresh-editor cold capture via an ADR-0009 poll job. Remaining limits include Epic MCP notifications/cancelled, durable-idempotency crash/migration caveats, intentionally partial graph/domain coverage, and metrics gaps. See docs/CAPABILITY_CATALOG.md and docs/proposals/ws-01-full-use-readiness-2026-07-30.md.

The problem

From the project owner, who has been living with the current tooling:

"We were having a lot of problems doing harder, more intricate things — generating effects and templates from the beginning, as well as micromanaging nodes on characters. Even using my own toolset I was getting five-to-one efficiency, but tool calls would still fail because we didn't have the base data, or it would try to set things before getting information from it."

Two distinct problems live in that sentence, and the second is the important one:

• Efficiency — too many round trips. Slow and expensive.
• Correctness — the agent acts without adjacent context, so calls fail, assets break, and repair loops burn more calls than the original work.

Because every tool call re-sends the whole conversation, cost is superlinear in call count and only linear in payload size. A 40 k payload sent once is cheap; a 1 k payload re-sent ninety-nine times is not. That inverts the usual instinct: return more context, in fewer calls.

Full reasoning and the cost model: docs/WHY.md.

The approach

1. Complete structured state. Whole graphs — Blueprint, Niagara, Material, AnimBP, Control Rig — retrieved and submitted as validated JSON, with diagnostics included by default rather than on request.
2. Verification before success. Nothing reports *_validated unless it was re-read and checked. "the tool returned OK" is not success.
3. Reusable patterns. A template library that composes and varies known-good constructions, so "a frost variant of the fireball" is one operation.

What already exists — and why it changes the design

UE 5.8 ships Epic's own agent stack, verified by source inspection (docs/GROUNDED_FACTS.md):

So we do not build an MCP server, a transport, a schema generator, or a rollback mechanism. We build the three things Epic did not: complete graph round-trip, verification, and the pattern library. See ADR-0002.

Repository layout

AGENTS.md              Operating contract. Every agent reads this first.
docs/
  WHY.md               The problem and the cost model. Read before designing.
  GROUNDED_FACTS.md    Verified UE 5.8 API surface, with source citations.
  adr/                 Frozen architecture decisions.
  WORK_ALLOCATION.md   15 workstreams, disjoint file ownership.
  research/            One brief per workstream.
  audit/               What Epic and REAgentTools already do.
  POC_ACCEPTANCE.md    Binary acceptance criteria.
  ROADMAP.md           Phased plan.
  RISK_REGISTER.md     What could sink this.
schemas/               Frozen JSON contracts: envelope, graph, batch, templates.
Plugins/UEREMCP/       C++ editor plugin (shipping on RE; build via UE 5.8).
tools/                 Schema validator, ownership guard.

Getting started

To launch the agent swarm: copy-paste prompts are in docs/SWARM_LAUNCH.md — one orchestrator, a worker template, and the four Wave 1 workstreams filled in.

If you are an agent working this repo: read AGENTS.md, then find your workstream in docs/WORK_ALLOCATION.md.

If you are a human: read docs/WHY.md, then docs/ROADMAP.md.

python tools/validate_schemas.py

Where to start work

Four Wave 1 questions gate everything else. All four are answerable in days, by opening a header or running code in the editor:

RB-05 is the one that decides whether the central thesis holds.

Related

• ../REAgentTools — prior art. 15 Python toolsets, ~5:1 measured efficiency. Its documented gaps are precisely this project's targets.
• $UEREMCP_LEGACY_PROJECT — the target project (UE 5.8).