Stefan VaskevichCan Claude Code actually drive Unreal Engine 5 - move objects, edit Blueprints, run the game, and look at the result? Yes. After a month of testing connectors and plugins (a lot of garbage, some of it expensive), the setup that genuinely works is two free, open-source plugins used together: UnrealClaude for viewport screenshots and moving actors, and VibeUE for Blueprint editing and Python. The only thing you pay for is Claude itself.
This is the full walkthrough: what each plugin does, how to install and connect them, and an honest stress test - I drop my AI-generated 3D assets into a project and ask Claude to build a playable endless runner from a third-person character, then judge how far it gets. Short version: it is a serious boost if you can describe the logic you want; vague prompts give vague results.
Watch the full build
The whole setup and stress test, start to finish, on YouTube.
1. The two free plugins that connect Claude Code to Unreal Engine
Almost every “AI for Unreal” plugin I tried either did not work or wanted money for the part that matters. These two are free, open-source, and cover different halves of the job - which is exactly why you run them together.
UnrealClaude - eyes and hands in the viewport
UnrealClaude is a plugin with 600+ GitHub stars that does the two things you most need: it can capture the viewport (so Claude can see what it just built) and it can move objects around the scene. It ships an MCP server with 20+ editor tools, targets Unreal Engine 5.7, and is MIT-licensed. That “take a screenshot and review yourself” loop is the single most useful thing in the whole setup.
VibeUE - Blueprints and Python
VibeUE is also open-source, but it pairs with its own dashboard, so it needs a free API key. You do not have to pay unless you want to use their own in-editor agent - we only want the MCP tools, and those are extensive: edit Blueprints, run Python scripts in the editor, work with assets and materials, and more. This is the half that actually changes your project.
2. Install and connect the MCP plugins
Prerequisite: install and open both Unreal Engine 5.7 and Claude Code, and open Claude Code in the folder of your Unreal project. From there the install is mostly conversational - you paste the two plugin links and tell Claude to handle it.
Drop the links, install both plugins
Give Claude the two repo links and say install these two plugins. The install pulls a few dependencies (Node.js, some Microsoft C++ libraries), so if anything is missing, just ask Claude to install it. You will also want a good CLAUDE.md in the project root - the instructions file Claude Code reads on every run. I wrote one by hand and you can grab it below; a proven setup file is where you save the most tokens and trial-and-error.
Get the free VibeUE API key
When the install finishes, Claude will flag that it needs the VibeUE key. Sign in at vibeue.com, copy the key from the dashboard, then paste it into the VibeUE settings (the gear icon) inside the Unreal Editor and save.
Install the MCPs and restart the editor
Claude may not register the MCP servers on the first pass - just ask it to install all the MCPs. Then restart the Unreal project so the freshly installed plugins load, and (sometimes) restart Claude Code so the MCP servers take effect.
Check both connections
Ask Claude to confirm both MCP servers are alive. UnrealClaude works independently with just your Claude auth; VibeUE only responds once the key is saved. When both reply, you are ready.
# A quick way to confirm both bridges are up:
curl http://localhost:3000/mcp/status # UnrealClaude
curl http://127.0.0.1:8088/mcp # VibeUE (405/200 = alive)
# In Claude Code, the simplest check is just:
# "check both mcp connections"Drop this into your Unreal project root as CLAUDE.md and Claude Code follows a known-good playbook - the .mcp.json wiring, the VibeUE skill library, and the ~20 hard-won UE 5.7 gotchas I hit so you do not have to (Blueprint edits locking during Play, the FBX-import crash fix, the silent button OnClicked no-op, and more).
3. The stress test: build an endless runner from prompts
I do not start from a blank scene - I use a third-person template with the default mannequin swapped for an AI-generated modular character (a fox, built in an earlier video; there is a full AI 3D character pipeline if you want to make your own). Then I just describe the game, one feature at a time, and commit after each one.
Clean up, then spawn an infinite path
First prompt: strip the accessories, keep only the character (Claude asks sensible questions like “keep lighting and sky?” - yes). Next: clear the level and create infinite tiles that spawn ahead and despawn behind as you move. It builds a BP_RunnerTile with readable variables, captures the viewport with UnrealClaude to confirm the character is standing on the planes, and it just works.
Auto-run, top-down camera, three lanes
Endless runners run on their own, so the next prompts make the character move forward automatically, pitch the camera down a bit, and add three lanes switched with A and D. Claude edits the existing Blueprints by executing Python scripts through VibeUE, then runs Play-In-Editor to test itself. The logic is good; the Blueprint layout is not.
Obstacles, coins, and a rising speed
Add obstacles you can crash into, then add coins as randomized collectibles. This round took about 15 minutes and ~14k tokens of Opus 4.8 - not free, but cheap for a working mechanic. I also had it ramp the run speed up over time, so the game gets harder the longer you last. One caveat: I never asked the coins to avoid the obstacles, so some spawn right on top of them - an easy follow-up prompt, and a good reminder that the agent does exactly what you ask, no more. Throughout, the same loop repeats: edit via Python, run the game, screenshot the viewport, read the result, fix.
Score, coins, game over + retry UI
A quick pass adds a score, a collected-coins counter, and a Game Over screen with a Retry button. The styling is rough - that is fine, it is a placeholder we can redesign later.
4. Swap the grey-boxes for AI-generated 3D assets
The mechanics work on grey-boxes; now make it look like a game. I generate a single concept image in ChatGPT, then extract the pieces I need from it - the obstacle, the coin, the bridge - and create the 3D environment separately. (Creating game-ready 3D models and environments is its own craft; I cover the free routes on my channel, and you can see how every generator ranks on the leaderboard.)
With the assets in, the requests are blunt: replace obstacles with the obstacle mesh, coins with the coin mesh, then build the bridge. I had a free toon shader from GitHub in the project that fought the new meshes, so I had Claude turn it off - editing materials is just another VibeUE use case. It was not bug-free: turning left or right falsely triggered Game Over, there were gaps between bridge sections, and obstacles clipped into the bridge railings. Most I fixed by showing Claude a screenshot and letting it work. The railing clip was the most interesting fix to watch: Claude fired a batch of line traces to map the bridge’s real geometry, then nudged the obstacles off the railings based on what the traces hit - measurement, not guesswork.
The final touches: project the background as a sphere/skybox (a faint seam remained from the source image - fixable), and add proper PBR maps to the bridge, which I had generated wrong the first time. Once the maps were right, it looked far better.
5. The result, and an honest verdict
Out came a genuinely playable endless runner: the fox auto-runs across a stone bridge through a floating-mountain world, dodges spiked rollers across three lanes, and the score climbs as it grabs coins. For a build driven almost entirely by prompts, with Claude doing the Blueprint and Python work, that is a real result.
- Give it detailed logic and architecture - describe how a feature should work and how it should scale - and you get amazing results.
- Give it vague, superficial prompts and it takes the fastest path, not the scalable one. Hard to build on later.
- It will not produce a pretty Blueprint. The logic works; the node graph is spaghetti and asking it to clean up mostly fails. Supervise it.
- Use Claude for organizing and logic, not asset generation or fine placement. Prepare those yourself.
Is this a full replacement for learning Unreal? Not quite - it is a massive boost for people who understand what they are asking for. If you are a beginner, do not avoid it; use it smarter. Ask it why a node is there, what a pattern does, and what the right architecture is for your game. As a learning partner on top of doing the work, AI is excellent.
/compact if you are staying on the same task - or start a fresh chat if you have moved to a new one. A new chat is cleaner than a compact; compacting tends to drag old noise along.What you actually get
That is the whole pipeline: connect Claude Code to Unreal Engine with two free plugins, prepare your AI assets, and let the agent assemble a game while you steer. New AI-3D tools that feed this kind of workflow land constantly - when they do, they go straight into the Arena so you can compare them before you commit. For more engine workflows, see building a UE5 level in a day and the no-code Aura workflow.
