An Epic Games Japan speaker showed at Unreal Fest Tokyo on 4 November how an AI coding agent built cutscenes and UI inside the Unreal Editor through Unreal MCP, and how much of the result depended on tools the team wrote around it. According to Game Makers’ session report, Suzuki (鈴木孝司) used Anthropic’s Claude models to assemble a MetaHuman conversation scene in Unreal Engine 5.8.3: speech-bubble UI, screen transitions, an auto-framing camera and Sequencer assets generated from a JSON storyboard. Everything was added as project code and plugins; the engine itself was not modified.
Unreal MCP is the Model Context Protocol server that Epic added to Unreal Engine 5.8 as an experimental feature, so that LLM agents can call editor functions. CGWorld described it in July, after an Epic Games Japan press roundtable, as a bridge between the engine and models such as Claude and Gemini. Suzuki’s talk, titled “How to use Unreal Engine MCP: working efficiently with LLMs”, was aimed at teams that have switched the plugin on and found that it is not enough. Game Makers headlined its report “Don’t stop at installing Unreal MCP”.
How the agent sees the editor
In the standard Tool Search mode the server exposes only three meta-tools, list_toolsets, describe_toolset and call_tool, so the agent loads a toolset’s description only when it needs it. Toolsets are written in C++ or Python, and their function names, argument types and code comments double as the documentation the model reads. Procedures and project-specific knowledge go into Markdown instructions and skills. Suzuki warned that context is limited, so more documentation is not automatically better.
Five tools the stock toolsets did not cover
- Sequencer events. The agent reported that it could not place event-trigger keys or create and bind endpoints, so a toolset was added that keys event tracks and wires Director Blueprint calls.
- Editor restarts. The agent had been killing the editor process, which risks corrupting assets. A script now checks for unsaved assets, closes the editor, builds, relaunches it and waits until MCP answers again.
- Screenshots of motion. Some shots finish within a second, so sequences are registered in advance and a subsystem pauses Play In Editor at the target moment for the agent to capture.
- UI animation. The agent first animated the 9-slice speech bubbles in Tick. A UMG animation toolset, modelled on the Sequencer tools, replaced that.
- Batch runs. The agent avoided
execute_tool_script, because a failed batch left it unclear how far it had got. A version that reports progress was added.
Not every refusal was real. The agent said it could not add a float track, yet the existing add_track_to_binding and set_property_name_and_path tools could do it in combination. “An AI answering ‘can’t’ does not necessarily mean the feature is missing,” the report notes. A second lesson was to build every tool as a set that can create, change and delete, and to test the whole sequence: deleting a UMG widget left orphaned animation rows behind, and recreating an animation with the same name crashed the editor.
Three cutscenes: 110 model calls down to 55
Suzuki timed the agent recreating three Sequencer cutscenes. Before tuning, Claude Sonnet took about 17.7 minutes, 110 model calls and 23.3 million cached-input reads. After tuning, it took about 10.8 minutes, 55 calls and 7.5 million reads. Claude Opus on the tuned setup took about 17.7 minutes and 73 calls; it checked save states and the screen more cautiously, with similar output quality for this workflow. The project started on Opus 5.0 and later used Opus 5.5 and Sonnet 5.5, at medium effort. Most of the time went on round trips between the model and MCP, not on editor processing.
The whole project consumed about 1.56 billion tokens. Roughly 40%, about 656 million, went into the session that turned the storyboard into Sequencer assets; rerunning the three cutscenes in a fresh session used about 24 million. Most of the volume was cached input, so multiplying the totals by list prices does not give the real cost, and the figures were not reconciled with actual billing.
Limits Suzuki listed
This was a verification project, not a commercial game in live operation. The camera rig does not cover all camera work. Sequencer binding and Blueprint-generation errors were only partly solved, and some fixes were made by hand. Long Blueprint graphs wired by the agent were hard for people to follow, and graphs with branches and return values produced compile errors. Local models were not tested, and neither was a comparison between MCP and plain Python batch scripts.
Screenshots came back to the agent as Base64, and a wrapper that returns a local file instead saved fewer tokens than expected. Returning images in MCP format is in a development branch and expected in Unreal Engine 5.9. At the same event’s keynote on 3 November, Tim Sweeney said a 5.9 preview build was due the following week, and that Unreal Engine 6 will drop Blueprints in favour of Verse from its early-access release in late 2027, with an AI conversion tool planned, Game Makers reported. The report on Suzuki’s session does not say whether the slides will be published.
For teams already on 5.8, the talk treats the shipped toolsets as a starting point: the measured gains came from project-specific C++ and Python tools, short instructions and timing each run. GameTales has previously covered Tokyo artist Hirokazu Yokohara’s agent-built Blender room ported to UE5, a solo experiment rather than an Epic workflow.
Disclosure: GameTales drafts its articles with AI assistance, including Anthropic’s Claude, the model family used in this talk.