4real plugs into Unreal 5.8's native MCP server and adds a deep editor toolset — terrain, UMG, Niagara, animation, materials, profiling — plus its own fleet of Claude Opus 4.8 agents that turn a natural-language goal into real editor actions.
Most MCP toolkits are hands without a brain — they wait for an external IDE agent to drive them. 4real brings its own. A lead orchestrator plans your goal into ordered steps and dispatches each to a domain specialist, all over MCP.
Landscapes, heightmaps, foliage, real-world/GPS terrain, PCG, map blockout.
UMG widgets and MVVM view-model bindings.
Niagara (incl. HLSL scratch pads), MetaSound, SoundCue.
AnimSequence keyframes, AnimBP state machines, montages, skeletons.
Graphs, enums/structs, data tables, gameplay tags, input, State Trees.
Materials, material graphs, landscape materials, UV mapping.
Frame timing, CPU/GPU-bound diagnosis, Insights traces, PIE.
Level actors, transforms, viewport, asset discovery.
A long tool-calling loop normally bloats and degrades a model's working set. The fleet stays sharp across big jobs with layered defenses:
memory_note / memory_recall findings instead of carrying them in-context.--verify) re-checks each step and retries gaps.4real complements Unreal 5.8's native toolsets — focusing on the depth the engine doesn't cover — and registers into the engine's own ToolsetRegistry, MCP server, and AgentSkill system.
Landscape sculpting, splines, auto-materials + RVT, foliage, FPS map blockout, real-world terrain from GPS.
Niagara with HLSL scratch pads, MetaSound & SoundCue graphs, AnimSequence/Montage/AnimBP, skeleton editing.
UMG + MVVM, higher-order Blueprint authoring — timelines, dispatchers, delegates, batch graph builder.
Frame timing with a CPU-vs-GPU-bound verdict, Unreal Insights trace capture, trace+log analysis.
Read-only scene_audit — actor/class counts, mesh & light totals, Nanite split, missing materials, as JSON.
Transaction service wraps undo/redo so an agent can group and roll back its own edits.
Drive the fleet from the CLI. --list-agents needs no API key or Unreal;
a full run needs an Anthropic API key, and a live editor needs the plugin built into a UE 5.8+ project.
Run the agents
# clone & install the agent runtime git clone https://github.com/ZubeidHendricks/4real cd 4real/Agents && pip install -r requirements.txt export ANTHROPIC_API_KEY=sk-ant-... # turn a goal into editor actions python -m fourreal_agents \ "Block out a foggy forest arena and add a HUD health bar" # or one specialist directly python -m fourreal_agents --agent profiler \ "Are we CPU- or GPU-bound this frame?" # inspect the fleet — no key, no Unreal needed python -m fourreal_agents --list-agents
Connect a live editor
# 1. enable Unreal's native MCP (Edit ▸ Plugins) # Unreal MCP + Editor Tools, then start the server # 2. drop 4real into your project cd YourProject/Plugins git clone https://github.com/ZubeidHendricks/4real ForReal # 3. point the fleet at the editor's MCP endpoint python -m fourreal_agents \ --transport http --url http://127.0.0.1:8000/mcp \ "Audit the level and tell me where the tris are"
Agents discover the editor's real tools via MCP
list_tools and filter them per specialist automatically.