Structured, part-aware 3D generation for AI agents. Named-part GLB, preview URL, Blender script.
io.github.RareSense/Nova3D MCP Server
The io.github.RareSense/Nova3D MCP server exposes Nova3D’s generation pipeline as a callable tool inside MCP-compatible AI agents. It supports “structured, part-aware 3D generation,” returning named parts in GLB format along with a preview URL, and uses a Blender script in the workflow.
🛠️ Key Features
Structured, part-aware 3D generation
Named-part GLB outputs (e.g., drum, door, control panel, hose connectors)
Preview URL provided
Blender script included
🚀 Use Cases
Generating editable 3D parts for AI agent workflows
Producing separate components instead of a fused 3D blob
Supporting tool-driven 3D generation in Codex, Cursor, VS Code, Visual Studio, Claude Code, and other MCP agents
⚡ Developer Benefits
One tool call to trigger Nova3D generation
Part-level outputs suitable for downstream editing (separately editable connectors/sections)
⚠️ Limitations
The available documentation excerpt describes setup and behavior, but does not specify tool name, tool count, or detailed input/output schemas beyond the named-part GLB and preview URL.
Robots harvesting on a Nova3D-generated farm, simulated in MuJoCo
A working farm, written by Nova3D from one prompt. Humanoids pick tomatoes, quadrupeds carry the harvest, drones scout. Simulated and filmed in MuJoCo.
Capabilities
Executable source: represent assets and environments as programs that can be inspected, edited, and rebuilt.
Parametric control: define dimensions, proportions, materials, physical properties, and motion through explicit parameters.
Controlled variation: specify which parameters change and which remain fixed across design alternatives or experiments.
Structural editing: change components, assemblies, layouts, clearances, and connections through their construction code.
Reusable construction: reuse components and generation logic to build related designs, asset families, and environments.
Geometry and motion checks: inspect dimensions, part relationships, pivots, and collisions, using checks appropriate to the design.
Versioned development: preserve source, parameters, build outputs, and validation results as the project evolves.
Target-specific exports: compile designs for CAD tools, 3D applications, and robotics simulators through dedicated export paths.
For robotics
Controlled scenario generation: specify which geometric, physical, and temporal parameters change, and which remain fixed.
Structural variation: change connectivity, obstacle arrangement, clearances, and mechanism configurations, not just textures or initial positions.
Reproducible evaluation: preserve the environment source, parameters, simulator configuration, and experiment results.
Failure-driven generation: eventually turn a failed rollout into a family of nearby scenarios to investigate the failure boundary.
Datasets
Our releases include Nova3D-Bench (preview), the benchmark used in our paper, and 20 outdoor simulation environments reconstructed from Wanderland captures, with editable scene source, annotations, and native simulator exports.
shooting_star Make this ring. → Turn it into a pair of matching earrings. → Make the earrings rose gold. → Change the earring shape into a rising-star climber.
People and AI agents work in the same world at the same time, each on their own branch. Then the branches merge into one world.
You: Build a cottage on the empty plot, with a glass tower. v1 · 53 parts
Maya: I want to see inside. Open up the back. v2 · 19 new parts inside
Sam: What does it look like in winter? v3 · snow, bare trees, frozen pond
Sam: Now make it night. v4 · 9 lamps and the fire lit
You: Merge Maya's open back with Sam's winter night. v5 · no conflicts
v6: the gardener bot waters the plants Maya: Send the gardener bot in to water the plants. v6 · in through the open back
worlds
Robots at work.
Every world below started as a text prompt, was generated as an executable scene program, compiled to MuJoCo, and populated with robots. Doors open on real hinges, produce carries labeled masses, floors carry measured friction. The environment is code, so every physical property is authored, not painted on.
image Café · Unitree G1 waiters clear tables; a Go2 serves from its back (learned walking policy, full physics)
image Mall · Spot patrols, H1 delivers parcels, Stretch restocks, a Skydio X2 flies the atrium
Farm · Apptronik Apollo picks tomatoes into a basket; ANYmal hauls the crate
Apartment · a G1 tidies a generated home, object by object
image Orbital lab · a UR5e throws a sample across the module in zero g; a Franka Panda catches it in flight
image Traction hall · one walking policy, three floors: ice μ 0.02, steel μ 0.30, rubber μ 1.05
Simulate. Test. Revise.
Run your designs in simulation to check movement, collisions, and interactions. Find what fails, revise the design, and test it again.
Test 1 · failed Husky, pick the ripe apples off this tree. The low limb stops it at 50 cm.
v2 · the low limb tied up 20° Test 2 · failed: the arm snags on branches, 0 of 3 apples picked.
v3 · 6 branches pruned out of the arm's path Test 3 · passed: 3 ripe apples in the bin, nothing knocked.
A Clearpath Husky with a UR10e arm and a Robotiq 2F-85 gripper, in an old orchard grown as code. Asked to adapt this tree for this robot and keep every apple, Nova revises the tree and tests again. Every move is the MuJoCo run: branches bend and spring back, stems snap only when pulled or lifted.
The real kitchen The scene was built from 8 phone photos.
Test 1 · failed Stretch, get the milk off the island. It bumps the bar stool, 1.1 m short.
v2 · adapted for Stretch, design kept Test 2 · passed: 4 things moved out of its path, milk in hand at 11 s.
physics
Physics, defined in source.
Mass, inertia, friction, and joint limits are explicit and editable alongside the geometry. Change the conditions and simulate the result. Below, the same scene runs with one parameter changed: gravity, from 0 to 9.81 m/s².
image g = 0 · a Booster T1 pushes off the hatch with its legs and floats the length of the module, carried only by the momentum of that push
image The ramp tilts slowly, and the blocks let go in order of grip: the ice block slides at just 4.6°, steel at 15°, aluminium at 20°, oak at 26°, and the rubber block holds on to the very end. Nobody animated any of this. Each block slips at exactly the angle where physics says its grip runs out.
Build on common ground.
Bring assets, scenes, and behavior together in one programmable project. Build from the same source for design tools, game engines, and simulators.