openPASO

open Platform for Agentic Simulation and Optimization

Describe a physics problem in plain words.
An AI model picks a solver, writes its input, runs it, and checks the answer.

_{Click the film to watch it in full quality on the documentation website.}

Warning

openPASO is under active development, and we invite you to help. It works and is used for real simulations, but it is young: things change quickly and you will find rough edges. Try it, tell us what broke, and share what you know about a solver. See Contribute.

What it does

Simulation programs, called solvers, compute how things physically behave: how a part bends, how heat spreads, how air flows. They are powerful and hard to use, and each one has its own input format and its own traps. openPASO puts nine of them behind one door and lets an AI model open it: you describe what you want in a normal sentence, and the model picks a solver, writes the input, runs it, and checks that the answer is computed correctly.

Quick start

You need Python 3.10–3.13 and about ten minutes. One solver is enough to begin.

pip install openpaso             # the server, with scikit-fem as a first solver
openpaso doctor                  # which solvers openPASO can use on this machine; no key, no network
openpaso                         # starts the MCP server on stdio; Ctrl-C to stop

Missing a solver? openpaso install ngsolve (or kratos, dune, ...) checks first and installs only what is missing, the same way the server's setup_backend tool would.

openpaso is the command your AI app's MCP configuration points at. Working from a checkout instead (to change the code, or to run check_install.py, which lists the solvers openPASO can use on your machine without a key or network):

git clone https://github.com/open-PASO/openPASO.git
cd openPASO
python3 -m venv .venv && source .venv/bin/activate
pip install -e . scikit-fem
python check_install.py

Then connect an AI model, one of two ways:

• Option A — an AI app you already have (Claude Code, Claude Desktop, Cursor): no extra cost. Set it up →
• Option B — your own OpenRouter key: one command per simulation, any model. Set it up →

Solvers

scikit-fem	NGSolve	Kratos Multiphysics
DUNE-fem	FEniCSx	deal.II
FEBio	4C Multiphysics	SPARTA

Documentation

Everything else is on the documentation website: installing each solver, both ways of use step by step, all 24 tools the model gets, how openPASO checks an answer, coupling two solvers on one problem, troubleshooting, and a glossary for every word.

Contribute

Reports of what did not work, solver traps you know, and plain-language fixes to the documentation are all very welcome. One rule stands above the rest: every improvement must help all simulations, not one example. See CONTRIBUTING.md and the Contribute page.

Licence and citation

MIT licence. openPASO was first published as OASiS; the archived releases and the paper use that name. If you use openPASO in research, please cite:

@software{openpaso2026,
  title  = {openPASO: an open-source multi-physics and multi-code framework
            for verified computer simulations},
  author = {Hermann, Alexander and Shojaei, Arman and Scheider, Ingo and Cyron, Christian},
  year   = {2026},
  doi    = {10.5281/zenodo.20543501},
  url    = {https://github.com/open-PASO/openPASO}
}

_{Helmholtz-Zentrum Hereon · Institute of Materials Mechanics  ·  Hamburg University of Technology}