LLM Agent in Fusion 360: Real-World Testing of Staircase Design Automation
The MCP LLM assistant in Fusion 360 enables generating 3D models from text prompts. Testing on a typical task—designing a wooden staircase—revealed both strengths and limitations. In just 3 hours, steps and risers were created, but the full project never came together. The Fusion API remains limited, making custom scripting difficult.
We started with basic treads: prompt "Create treads 300 mm wide, 900 mm long, 40 mm thick. Overhang 50 mm, rise height 190 mm, first step 135 mm, 12 treads."
Results:
- Geometry matches specifications.
- Issues: excessive objects in the timeline, independent sketches lacking parametric links.
Adding Risers and Collision Problems
Risers are vertical boards matching the rise height (190 mm), 900 mm wide, and 16 mm thick. The initial prompt caused intersections with treads.
Fine-tuning required precise offsets: "Move risers 20 mm along Z, 8 mm along X." Outcome: near-perfect positioning, but no proper connections or dependencies. The model loses dynamic behavior—changing base parameters no longer propagates through the design.
Parametric Challenges with Newel Posts
Attempted modeling of two 80×80 mm newel posts beneath the final tread. Instructions: top of posts aligned to bottom face of tread, 20 mm offset from edges, height extending to floor plane (Plane72).
The agent spent 5 minutes processing and returned posts sized incorrectly—starting from the first tread. Recommendation: always use version control before critical stages.
Second attempt: one post from Plane73 to Plane72, with 20 mm offset from Plane74 and 0 mm from Plane75. After 15 minutes, the agent kept shifting reference planes and generating clutter. The project collapsed.
Third try: manual sketch for extrusion. The agent requested profiles but failed due to ambiguity in the sketch.
Additional limitation: the agent doesn’t support parallel work across multiple projects—it halts when switching.
Key Agent Limitations
Testing exposed workflow bottlenecks:
- No image or sketch attachment for clarification.
- No object/face/profile selection within the model context.
- Single chat only: can’t test multiple variants simultaneously.
- No quick prompt editing (Ctrl+Shift+arrows).
- Relies on rigid coordinates instead of constraints.
- Timeline gets cluttered with unnecessary sketches and planes.
| Aspect | Status | Impact |
|--------|--------|--------|
| Parametric Links | Missing | No model dynamism |
| Object Selection | Not available | Prompts become overly verbose |
| Timeline Clutter | High | Manual cleanup required |
| Parallel Projects | Impossible | Slows iteration cycle |
What Matters Most
- LLMs speed up prototyping simple shapes but fail at dependency management.
- Useful as a rough draft for mid-to-senior CAD engineers—but requires manual refinement.
- Fusion API remains poorly documented; scripts function only in restricted modes.
- Potential lies in multi-chat workflows and visual input for production use.
- 3 hours for partial staircase vs. same time spent manually designing it.
Overall, the MCP LLM is a proof-of-concept for generative design in Fusion 360. With improvements in parametric logic and UI, it could become competitive.
— Editorial Team
No comments yet.