mcp-fea

Real finite element analysis from a conversation. An AI assistant sends a STEP file and a plain-English question ("does this arm survive a 6 g maneuver with a 400 g motor?"); this MCP server maps the words onto geometry, rejects bad setups deterministically before any compute is spent, meshes and solves with CalculiX on Modal as an async MCP task, and returns a verdict backed by numbers — peak von Mises stress, factor of safety, deflection, natural frequencies — plus rendered stress plots and an interactive 3D viewer. Every study ends with a non-skippable reaction-balance check; every benchmark on the validation page is regenerated by CI on every commit.

Measured on the live deployment (2026-07-31): cold start 3.8 s · cantilever study submit→completed 9.3 s warm / 12.9 s with a cold solver container · ~$0.002 per study at Modal's posted rates.

Full table, mesh stats, and an explicit limitations section: docs/validation.md — regenerated by CI, never hand-edited.

Protocol: MCP 2026-07-28 (stateless core) with the io.modelcontextprotocol/tasks extension implemented per SEP-2663 (the official Python SDK v2 shipped without it — this repo also contains the client-side extension that teaches the official SDK to consume task-augmented servers) and an MCP Apps (SEP-1865) 3D viewer. See docs/decisions.md for every engineering call and why.

Deploy (3 commands)

uv sync
uv run modal secret create mcp-fea-token MCP_FEA_TOKEN=$(openssl rand -hex 24)
uv run modal deploy -m src.server.modal_app

Modal prints the endpoint URL (https://<workspace>--mcp-fea-mcp-endpoint.modal.run). Connect any 2026-07-28 MCP client with Authorization: Bearer <your token>. First-time Modal users: uv run modal token new once. Then verify the live deployment end to end:

MCP_FEA_LIVE_URL=https://... MCP_FEA_TOKEN=... make verify-live

Tools

Task lifecycle: working (statusMessage walks validating → meshing → solving → postprocessing) → completed | failed | cancelled; tasks/cancel kills the running container. Tasks are retained 24 h.

Units policy

One conversion boundary, at ingest. Internally everything is consistent mm–N–MPa(–tonne): lengths mm, forces N, stress/moduli MPa, density tonne/mm³ (kg/m³ × 1e-12 — done once, in Material.density_tonne_mm3), which makes modal frequencies come out directly in Hz. STEP coordinates are interpreted in the declared units (mm|m|in); gmsh's own unit conversion is disabled so the declared unit is authoritative — and sanity-checked (a 100 mm part declared "m" gets flagged).

Validation codes

Every rejection returns {code, message, suggested_fix}.

Abuse limits (all env-tunable)

25 MB STEP cap · 40 MB request cap (413) · per-token rate limit (MCP_FEA_RATE_LIMIT/_WINDOW_S, 429) · concurrent-solve cap (MCP_FEA_MAX_CONCURRENT) · daily budget (MCP_FEA_MAX_SPAWNS_PER_DAY) · kill switch (MCP_FEA_DISABLE_SPAWN=1). CAD parsing never runs in the endpoint process (see SECURITY.md).

Local development

uv sync
curl -Ls https://micro.mamba.pm/api/micromamba/osx-arm64/latest | tar -xj bin/micromamba
./bin/micromamba create -y -p .tools/ccx-env -c conda-forge calculix=2.23
make test        # fast suite (protocol, benchmarks, validation, hardening)
make torture     # 21 real-world STEP files through the endpoint
make validation-page
uv run uvicorn --factory src.server.app:create_local_app --port 8000  # local server

Repo map

src/pipeline/ geometry, meshing, deck, solver, validation, post, render · src/materials/ library · src/server/ protocol, tools, tasks, limits, worker host, Modal app, Apps UI · src/client/ SDK tasks extension · tests/ (incl. torture/) · demo/ · docs/ decisions, validation, upstream notes. License: Apache-2.0.