3D Agent — it plans first, and never runs on its own

Open 3D Agent →

3D Agent turns a sentence into a named plan: at most six steps, each with a label, a one-line reason and the exact parameters it would use. It does not execute. Nothing happens until you press Run all steps, and then the whole chain runs in one session with each step landing as a node you can open in 3D Studio. Planning costs nothing — the credits go to the work, not to the conversation about it.

What 3D Agent makes

The plan is the first artefact and the point of the app: you can read exactly what would be spent before anything is. Run it and the second artefact appears — a chain of nodes, one per step, in a single session.

There are three ways to start a plan: from an idea in words, from one to four photographs, or from a .glb model you already have.

Walkthrough: from a sentence to a rigged model

1

Say what you want

A message up to 2,000 characters. "A low-poly market stall, game-ready, rigged for an idle animation, exported as GLB" is the shape that works — an outcome and its constraints, not a list of tool names.

2

Or start from what you have

Attach one to four photographs of an object, or import a .glb. Importing a model costs nothing and accepts files up to 100 MB.

3

Read the plan

Each step carries a label, one line of why, and the exact parameters. This is where to catch a face limit that is too aggressive or a rig that is not the motion you meant — changing your mind here is free.

4

Press Run all steps

The chain executes in one session. Each step becomes a node; open any of them in 3D Studio to inspect or keep working.

5

Close the tab if you like

Steps store a request id, so reopening the tab re-attaches to jobs that are still running rather than losing them.

The seven tools it can plan with

  • generate_3d_from_text — a model from a description.
  • generate_3d_from_image — a model from your photographs.
  • remesh — low-poly, game or quad topology, with a face limit from 1,000 to 200,000 (default 20,000).
  • retexture — new surfacing on an existing mesh.
  • repair — watertight or hollow.
  • rig — none, walk, run, idle or jump.
  • export — glb, obj, stl, fbx, usdz or 3mf.

Limits

  • 6 steps per plan. A plan you cannot read is a plan you cannot approve.
  • 2,000 characters per message.
  • 1 to 4 photographs when starting from images.
  • 100 MB for an imported model, and importing is free.
  • A plain question returns zero steps — the agent answers it instead of inventing work.
  • If the planning model is unavailable, a deterministic keyword planner answers instead, so the app degrades rather than breaking.

What it costs

3D Agent is a Pro app, and the split inside it is unusually clean:

  • Planning is free. No credits, however many times you re-plan.
  • Running charges per vendor operation — the same prices the equivalent 3D Studio action would cost.
  • Importing a model charges nothing.

That is why the plan shows its parameters: you are being asked to approve a bill, so you are shown the bill.

Frequently asked questions

Will it start generating as soon as I describe something?

No, and this is the app's main design decision. A chat turn produces a plan and nothing else. Execution only happens when you press Run all steps.

Does planning cost credits?

No. Plan, re-plan and argue with it as much as you like — credits are spent by running steps, not by writing them.

Can I change a step before running it?

Ask for the change in words and re-plan. The new plan shows the new parameters, still for free.

What if I just ask a question?

You get an answer and zero steps. The agent does not manufacture a build out of a question about one.

What happens if I close the tab mid-run?

Each step stores a request id, so reopening 3D Agent re-attaches to the jobs still running.

Where do the results end up?

As nodes in one session. Open any of them in 3D Studio to inspect, edit or export further.