Agentic FEA: Parametric Motor Mount via Claude + MCP + CalculiX

Claude drives a full open-source FEA loop through the Model Context Protocol: Gmsh (CAD + 10-node tet mesh) → Abaqus-format .inp → CalculiX → parsed results → plots. No commercial licence needed. Ask in plain language, e.g. "Make the wall thinner until f1 drops below 1100 Hz, then tell me the lightest design that keeps safety factor > 5 at 60 N thrust."

Part: fixed-wing / VTOL motor mount (L-bracket + gusset, 4× motor bolts on a 19 mm circle, 4× base bolts), Al 6061-T6. Analyses: modal (motor mass lumped on bolt nodes) and static (thrust + reaction torque), parameter sweep, mesh convergence.

Results

Mesh convergence (20 N thrust, motor mass 60 g):

Related MCP server: mcp-fea

Quick start

sudo apt install calculix-ccx          # or set CCX=/path/to/ccx
pip install -r requirements.txt
python scripts/validate_cantilever.py  # solver-chain check vs beam theory
python scripts/run_demo.py             # build, modal, static, plots in assets/
pytest

Claude Code: the repo ships .mcp.json. Claude Desktop: add {"command": "python", "args": ["/abs/path/mcp_server/server.py"]} under mcpServers.

MCP tools

set_design · get_design · run_modal · run_static · plot_mode_shape · plot_stress_field · parameter_sweep · mesh_convergence

Known limitations:

• Peak von Mises at the fixed base-bolt holes keeps rising with refinement (6.4 → 9.7 MPa). That is a stress singularity from fully-fixed hole walls, not a converged value. Use displacement, frequency and stress trends; for real stress, model washer-face contact or bolt preload.

• Gmsh meshing is non-deterministic. On rare runs, a poorly-shaped element near the singularity can produce a spurious peak VM one to two orders of magnitude above the typical value (e.g. 196 MPa vs ~8 MPa). Re-running regenerates a clean mesh. The convergence table above reflects a clean run.

Modelling assumptions

Units mm-t-N-s-MPa · base-hole walls fully fixed · motor = lumped mass spread over motor-bolt nodes (no rigid coupling) · linear elastic, small displacement · no fillets.

Abaqus

The .inp uses Abaqus keywords (*NODE, *ELEMENT C3D10, *FREQUENCY, *STATIC, *CLOAD, *NODE PRINT), so fea/inp.py is the porting path. An Abaqus runner and .dat parser are not implemented or tested; fea/solver.py is the one place to add them.

Layout

fea/geometry.py  parametric CAD + mesh      fea/inp.py     .inp writer
fea/solver.py    ccx runner + .dat parser   fea/pipeline.py  MountModel API
fea/plots.py     3D mode/stress plots       mcp_server/server.py  MCP tools
scripts/         demo + validation          tests/         pytest

Credits

Inspired by vamshidharre/Modal_Analysis.