Skip to content

FAQ

Weld is in private beta and invite-only. A project owner on a team’s hub invites you by email; your inviter sends you an install link: the hub’s welcome email has the install steps, and so does the Install Weld page on your hub once you’re signed in. These public docs don’t include installers. Not on a team yet? Request access.

No. A hub only accepts people a project owner invited (weld invite <email>). Joining with an address that has no invite gives WELD-E2032.

Two or more people, each with Claude Code on their own account, and an Unreal Engine 5.8 project (or start one with weld game new). Macs (Apple silicon) and Windows PCs both work as nodes. The asset pipeline needs an NVIDIA GPU with 12 GB or more of VRAM somewhere on the team.

Unreal Engine 5 with C++ and Blueprints, including Epic’s official Unreal MCP server. Weld’s engine layer is built so Unity and Godot can follow.

Only one machine on the team needs one, for concept images and image-to-3D. Everyone else sends it jobs. Builds and smoke tests need Unreal, not a big GPU.

On your team’s hub: it hosts the game’s git repository and Git LFS objects, Work items, assets and telemetry. Agents run on your own machines under your own Claude accounts.

What does Weld send to the hub from Claude Code?

Section titled “What does Weld send to the hub from Claude Code?”

Metrics and costs (tokens, cost, errors) always, so the team can see spend per person and agent. Your prompt text and tool details only if you agree when joining (weld node telemetry --no-prompts turns it off at any time). Raw tool output is not captured. Diagnostic bundles are redacted (no secrets, no prompt text) and only written when you ask.

Only what you allow. Typed jobs (builds, smoke tests, asset jobs) from teammates whose keys you trusted after comparing fingerprints run automatically; shell jobs need your approval in your own terminal; agent runs never leave the submitter’s own machines. See Jobs and trust.

Does my Claude subscription get used for someone else’s work?

Section titled “Does my Claude subscription get used for someone else’s work?”

No. AI work never crosses subscriptions. A handoff passes the task; the recipient’s own agent does the work. The merge agent runs under the account of the person who is merging, and background agents under the node owner’s account, for their own work only.

Never. Anyone can edit any file. Weld prevents feature collisions by making Work visible, asking once about overlaps and predicting conflicts exactly with git.

What happens to binary assets like Blueprints and maps?

Section titled “What happens to binary assets like Blueprints and maps?”

They can’t be merged line by line, so Weld warns early when two branches touch the same one, keeps logic in C++ and maps in One File Per Actor to make it rare, and, if it happens, shows both sides’ op logs and keeps both versions while a human picks one.

No. Hooks fail open: if weld or the hub has a problem, Claude keeps working and shows a one-line notice with an error code. The only thing that can pause an edit is the one-time overlap question, and only once per overlap.

Keep working. git, Unreal and Claude Code work without it, Work updates queue locally and reconcile on reconnect. Coordination, the merge queue and remote jobs pause until it’s back.

Only the very first push of a repo. After that, changes land through the merge queue (weld merge); the hub refuses direct pushes to main.

Small teams: two to a handful of people, each running one or more agents.

Pricing isn’t set yet; beta teams hear about it before anything changes. Your agents run on your own Claude Code subscriptions, so Weld never resells model usage. weld ops costs shows Claude spend per person, agent, model or machine.