Here’s a session that took about four minutes. No blueprint nodes. No drag-and-drop. No clicking through panels. Just a terminal and a prompt.

The setup: Claude Code connected to Webified Bridge, UE5 editor open on an empty level.

The prompt

Spawn five SM_Rock assets in a 200cm grid at origin, add a three-point lighting rig above them, frame the viewport on the selection, and take a screenshot.

That’s it. One sentence. Here’s what happened next.

What Claude actually did

Webified Bridge exposes 151 tools to Claude Code. Claude picked the right ones, in the right order, without being told how:

→ level.find_asset_by_name       "SM_Rock"
→ level.spawn_actor_from_class   x5  (grid positions calculated)
→ recipe.lighting_rig            "three_point" above selection
→ viewport.frame_selection
→ editor.screenshot              path="session_001.png"

Sixteen tool calls. About 40 seconds. The Output Log confirmed each one as it fired.

The result

Five rocks in a clean grid. Three lights — key, fill, rim — positioned automatically above the selection. Viewport framed. Screenshot saved to disk.

The Outliner went from empty to this:

SM_Rock_0   (0, 0, 0)
SM_Rock_1   (200, 0, 0)
SM_Rock_2   (400, 0, 0)
SM_Rock_3   (0, 200, 0)
SM_Rock_4   (200, 200, 0)
PointLight_Key
PointLight_Fill
PointLight_Rim

No mouse. No editor panel. The agent read the scene state after each step and adjusted — it checked that the assets existed before spawning, verified actor placement before framing, confirmed the screenshot path before saving.

Then I asked it to clean up

Delete everything you just spawned.

This is where the approval gate kicks in. level.destroy_actor is a destructive tool — it requires approved: true before it runs. Claude surfaced the list and asked:

I'll delete these 8 actors:
  SM_Rock_0, SM_Rock_1, SM_Rock_2, SM_Rock_3, SM_Rock_4,
  PointLight_Key, PointLight_Fill, PointLight_Rim

Shall I proceed?

I typed yes. Eight actors removed. Level back to empty.

Nothing gets deleted without a confirmation step. That’s not a workaround — it’s by design. Destructive tools in Webified Bridge require explicit sign-off.

Why this matters

The interesting thing isn’t that Claude can spawn rocks. It’s that the agent maintained context across 16 tool calls, used the results of earlier calls to inform later ones, and stopped to ask before doing something irreversible.

That’s the difference between a copilot that autocompletes inside the editor and a control plane that runs the editor as a structured API. Webified Bridge is the second thing.

Try it

If you’re already on Claude Code, setup is one config block. The sidecar handles the rest.