opennodex

Structural finite-element analysis and structural health monitoring from .ndx decks — from Python, the command line, and any MCP-capable agent harness.

pip install opennodex (Python 3.10+) brings everything a run needs: the .ndx compiler, the native nodex engine and the native nodex-mesher, built for your platform; the nodex-shm SHM engine (operational modal analysis, damage localization, monitoring) with its numpy/scipy/pandas/matplotlib; and Node.js as a dependency wheel. Nothing else to install, no solver to configure.

pip install opennodex
opennodex selftest          # a beam, an auto-meshed slab, the slab over MCP, and an SHM analysis

Command line

opennodex check viga.ndx               # compile only: positioned diagnostics
opennodex run viga.ndx                 # compile + solve, one summary line per solve
opennodex run viga.ndx --json r.json   # …and the full result
opennodex info                         # compiler, engine contract version, mesher

Python

import opennodex

out = opennodex.run("viga.ndx")        # a path, or the deck's text
for solve, result in zip(out["solves"], out["results"]):
    print(solve["kind"], len(result["nodes"]))

opennodex.check("viga.ndx")            # {"ok", "diagnostics", ...} — never raises for a bad deck

MCP — the opennodex tools in any agent harness

opennodex mcp is a local MCP server (stdio) with the opennodex agent's tools: validate, run (native engine, real runId), query results, check design criteria, section catalog, suggest a section, compare modal data, twin training, damage localization and mechanism diagnosis (nodex-shm), generate a frame, import SAP2000 / ETABS / Abaqus / OpenSees / SOFiSTiK models, bake a parametric deck, model patches and more.

It never edits your files: your agent does, with its own tools, as diffs you see. The generators (generate_frame, import_model, bake_deck) return the .ndx they would write, for your agent to save. Your agent also gets the opennodex agent's own instructions — its FEM rules and the whole .ndx reference — through the server's instructions and the opennodex_instructions tool (some hosts shorten or drop server instructions; the tool is never shortened). File paths resolve against the folder the host starts it in (your project), or --root DIR.

Install it once so opennodex is on your PATH (pipx install opennodex or uv tool install opennodex), or let uvx fetch it on demand ("command": "uvx", "args": ["opennodex", "mcp"]). Then register it:

Claude Code

claude mcp add opennodex -- opennodex mcp

opencode — opencode.json

{ "mcp": { "opennodex": { "type": "local", "command": ["opennodex", "mcp"], "enabled": true } } }

OpenAI Codex CLI — ~/.codex/config.toml

[mcp_servers.opennodex]
command = "opennodex"
args = ["mcp"]

Claude Desktop / Cursor — claude_desktop_config.json / .cursor/mcp.json

{ "mcpServers": { "opennodex": { "command": "opennodex", "args": ["mcp"] } } }

VS Code — .vscode/mcp.json

{ "servers": { "opennodex": { "type": "stdio", "command": "opennodex", "args": ["mcp"] } } }

opennodex selftest also drives the MCP server the way a host does and runs a deck through it.

A deck looks like this:

model units kN, m
material steel E=210e6 nu=0.3 rho=7.85
section s A=0.00538 Iy=8.36e-5 Iz=6.04e-6 J=2.01e-7
node A at (0,0,0)
node B at (6,0,0)
beam B1 from A to B steel s div 4
fix A
load 1 nodal B fz=-10
solve linear_static

Every number comes from a real engine run. The deck carries its own section properties, so it runs the same on any machine.

Platforms

Linux x86_64 (glibc as tagged on the wheel), Windows x64, macOS (Apple Silicon and Intel). The desktop app (editor, 3D viewer and the optional agent) is a separate download: opennodex releases.

Español. pip install opennodex instala el compilador de .ndx, el motor nativo nodex, el mallador nativo y el motor SHM nodex-shm para tu plataforma, más Node.js. opennodex selftest comprueba la instalación; opennodex run deck.ndx compila y resuelve; opennodex mcp expone las herramientas de opennodex como servidor MCP local para Claude Code, opencode, Codex, Cursor o VS Code (configuración arriba). La aplicación de escritorio se descarga aparte.