Weld coordinates your team and its Claude Code agents on one Unreal project. Overlaps get caught before anyone edits, conflicts are predicted, and a merge agent lands the work.
The problem
Agents move fast and never look up. They don’t know what your teammate’s agents are changing, and you find out when the build breaks. Weld is the layer between them.
How it works
Everyone keeps working in Claude Code. A small node on each machine connects the agents to your team’s hub over MCP and hooks; the hub coordinates the work and sends jobs to whichever machine fits.
Each agent declares its feature and scopes before it edits.
Overlaps and predicted conflicts reach the right human, once.
Critics gate quality; the queue and merge agent land the work.
Build and smoke-test on the Windows PC, then upload to itch.
01 · Overlaps
Agents declare feature intents, not file locks. When two pieces of work share a system, each agent asks its human one question: coordinate, proceed or pivot. The answer reaches the other side.
02 · Conflicts
git merge-tree runs against live branches, so you see a shared file while there is still time to talk. At merge time, an AI merge agent resolves text conflicts with both intents in hand, and a person approves.
03 · Quality
Fresh-context critics that never see the worker’s reasoning score each iteration against your team’s rubric. Below target, the merge waits for a person.
Ember lantern prop: art 7.4 → 8.7 in three iterations, past the 8.5 target.
One dashboard
Merges waiting on you, critic gates, every agent’s live activity and which machines are free.
AI asset pipeline
Ask an agent for a charred iron lantern. Weld runs the concept, generates the mesh on the team’s GPU, cleans it up in headless Blender and imports it into Unreal through Epic’s official MCP. A person approves each step.
Every asset records its generator, model, license, prompt and job. A license gate holds any merge that uses a model outside your project’s allowlist until a person signs off.
SM_Lantern
Content/Props/Lantern/SM_Lantern.uasset
Lantern 
Brazier 
Iron chest 
Crawler 
Mushrooms 
Barrel 
Sword 
Torch Built for Unreal and Claude Code
> add drag-and-drop to the inventory grid
⏺ weld_work_declare("Inventory grid UI", scopes: [system:inventory, ui-screen:inventory])
⎿ overlaps Maya’s “Loot drop + pickup” on system:inventory
⏺ Maya’s agent is adding pickup to the same inventory system. Coordinate, proceed or pivot?
> coordinate. I’ll stay in the grid widget
⏺ weld_overlap_ack("coordinate")
⎿ Maya chose coordinate: “I only need UInventoryComponent::AddItem.”
04 · Operations
Health, logs and Claude costs for every machine, over OpenTelemetry. Every error says what happened, why, and how to fix it; a doctor repairs what it safely can, and updates roll back on failure.
Also in the box
Your hub serves git and Git LFS for the game. One URL, one login, no lockable files.
Typed, signed jobs run where they fit: packaging on the Windows PC, generation on the GPU box.
Package, smoke-test headless with screenshots, upload. One step from green build to playable page.
Hand a task to a teammate or their agent, message people, and see what needs you in one inbox.
People sign in with passkeys. Machines and agents get scoped device tokens; secrets lease per job.
Claude tokens and spend per person and per agent, straight from Claude Code’s OpenTelemetry.
Built with Weld
151 s
Our jam test project, packaged as a Win64 build on the Windows PC.
9 s
FLUX.1-schnell on an RTX 4080 SUPER, the first step of the asset pipeline.
76 s
TRELLIS.2 at 1024 cascade on the same card, image in, textured mesh out.
Single runs on our hardware in October 2026. Your timings will vary with project size and GPU.
We built Weld for ourselves.
We make Unreal games mostly by driving Claude Code agents: two people in different places, each with a Mac and a Windows gaming PC. Without coordination, agents on the same game walk into each other, and nobody wants to be on a call all day to prevent it. Weld is how we work now, and we are opening it to a few more teams.
Weld is invite-only while we work closely with a small number of teams. Request access and tell us what you are building. We send invites in small batches and help each team get set up.
Unreal Engine 5 with C++ and Blueprints comes first, including Epic’s official Unreal MCP. Weld’s engine layer is built so Unity and Godot can follow; tell us on the form if that is what you use.
Two or more people, each with Claude Code on their own account, and an Unreal project. Mac and Windows machines both work. The AI asset pipeline needs an NVIDIA GPU with 12 GB or more of VRAM somewhere on the team; everyone else sends it jobs.
Pricing is not set yet. Beta teams will hear about it before anything changes. Your agents run on your own Claude Code subscriptions, so Weld never resells model usage.
On your hub. Weld hosts your game’s git and LFS on a hub your team runs, on a spare machine or a small VPS. Agents run on your machines under your own Claude accounts. Code, assets and telemetry stay on your hub.
Small teams: two to a handful of people, each running one or more agents. That is the setup Weld is built and tested for every day.
Private beta
We are inviting small teams building with Unreal and Claude Code. Tell us about yours.
Request access