souxmar

An open-source CAE platform: parametric CAD, mesh generation, FEM and CFD, post-processing — wrapped in a cross-platform desktop app with an agentic AI chat that can drive the entire pipeline. C++20 core, Python bindings, stable C plugin ABI, Tauri + React desktop app, Apache 2.0.

_{Above: the Rosenthal temperature field and melt-pool isotherm computed by the in-tree am-thermal plugin at its default 316L / 200 W / 0.8 m/s parameters. Regenerate it with python3 scripts/gen-physics-figures.py.}

For mechanical, structural, aerospace, manufacturing and marine engineers who want a Cursor-style experience for simulation work: open the app, describe the problem in chat, watch it mesh and solve, inspect results in a built-in viewport.

The strongest part of the repository is the physics. docs/PHYSICS.md documents every model in it — the governing equation, the literature citation, the validity envelope, and the direction and magnitude of its known error. If you only read one page, read that one.

Documentation site: https://celikgo.github.io/souxmar/ — the same docs rendered and searchable. CI fails if it stops resolving.

`v0.9.0`

What this is NOT

souxmar does not replace FreeCAD, Gmsh, FEniCSx, OpenFOAM, Blender or ParaView. It unifies them under a shared data model and a stable plugin ABI, then puts a modern UI and an agentic AI on top.

More specifically, and more importantly:

• There is no CAD kernel. include/souxmar-c/brep.h and sketch.h are ABI surface only; the in-core backing returns NOT_IMPLEMENTED from every operation and no cad.* plugin exists here. Parametric modelling, the feature tree and the 2D sketcher are designed (RFC-0003, RFC-0005) and unbuilt.
• The always-on solvers are demonstration stubs. solver.heat.linear, solver.elasticity.linear, solver.modal.linear and solver.cfd.simple return closed-form fields so that the pipeline, the examples and the agent evals have something runnable in the default CI matrix. They are not FEM and not CFD — no stiffness matrix is ever assembled. The one real discretised solve, solver.heat.fenicsx (DOLFINx + PETSc), is behind an opt-in build flag.
• The manufacturing and marine models are closed-form and heuristic. Every one cites its source and states what it ignores. They are screening and preliminary-sizing aids, not calibrated process simulations. The marine qualification dossier is advisory: souxmar is not a classification society and issues no approval.
• The viewport does not render yet. The desktop app's 3D panel is scaffolded; the renderer behind it is RFC-0001 and unbuilt.
• The Pro-tier services are scaffolds, not deployments. services/ contains API shapes and handlers with nothing running behind them. Every *.souxmar.invalid hostname in this repository is a placeholder on the RFC 6761 reserved TLD, chosen so it cannot be mistaken for a live endpoint.

docs/CAPABILITIES.md states this per capability, and is generated from the tree rather than written by hand.

Status

Pre-1.0, single maintainer, actively developed. v0.9.0 is the first tagged release; everything below is checkable from a clone.

Version	v0.9.0 — see VERSION, CHANGELOG.md
Licence	Apache-2.0
Plugin C ABI	v1 major frozen (ADR-0008); minor at v1.9, ratcheted additively
Agent tool contract	v1 frozen (ADR-0011); 24 tools
In-tree plugins	25, providing 38 capabilities
Tests	740 gtest cases across 67 files
C/C++	~66,800 tracked lines
Design record	45 ADRs, 9 RFCs
CI	CI, Security and Visual regression workflows on every push

What is not true, and has been claimed here before: there is no stable 1.0, no PyPI package, no hosted service, and no team. If you find a statement on this page you cannot verify with git, gh and a browser, that is a bug — please open an issue.

60-second quickstart

git clone https://github.com/celikgo/souxmar.git
cd souxmar
export VCPKG_ROOT="$HOME/vcpkg"        # https://github.com/microsoft/vcpkg

cmake --preset dev && cmake --build --preset dev

Then run a real pipeline — read an STL, compute per-cell mesh quality, write a ParaView file:

cd examples/stl-cube
../../build/dev/src/cli/souxmar run pipeline.yaml \
  --plugin-path ../../build/dev/examples/plugins
# -> cube.vtu

The manufacturing chain end to end — layered build mesh, LPBF thermal history, melt-pool porosity risk, inherent-strain distortion, residual stress, overhang check, three ParaView files and two Markdown build reports:

souxmar run examples/am-lpbf-bracket/pipeline.yaml \
  --plugin-path build/dev/examples/plugins

Read examples/am-lpbf-bracket/README.md first — it is explicit about what that pipeline does and does not model.

Surfaces

• Desktop app (macOS / Windows / Linux) — chat panel, pipeline editor, inspector, and a viewport panel whose renderer is not built yet.
• CLI (souxmar) — for CI, batch runs, scripting. This is the surface that works today. v0.9.0 publishes unsigned CLI tarballs for linux-x64, macos-arm64 and windows-x64, but they ship no plugins and no plugin SDK headers — and every capability lives in a plugin, so a downloaded build cannot run a pipeline. Build from source.
• Python (pysouxmar) — not on PyPI. pip install pysouxmar does not work and never has. Build the bindings from source with the dev-python preset:

  cmake --preset dev-python && cmake --build --preset dev-python

• Plugin SDK — stable C ABI for shipping your own meshers, solvers, elements, readers and writers as out-of-tree binaries. See docs/PLUGIN_SDK.md.

Capabilities

The full table — every capability id, whether it is implemented, tested, a stub, or planned, what kind of model sits behind it, and which test covers it — is in docs/CAPABILITIES.md. That file is generated by scripts/gen-capability-table.py from the plugin manifests, the CMake build wiring and the test sources, and CI fails if it drifts.

Summary of the 25 in-tree plugins:

Plugin	Capabilities	Notes
hello-mesher, hello-writer	mesher.tetra.hello, writer.text-summary	Minimal ABI references
grid-mesher	mesher.tetra.grid	Structured grid over the bounding box
gmsh-mesher	mesher.tetra.gmsh	Opt-in, real conforming mesher
stl-reader, obj-reader	reader.stl, reader.obj	Always-on surface readers
occt-reader	reader.step, reader.iges	Opt-in, OpenCASCADE
blender-reader	reader.blend	Opt-in, Blender subprocess
vtu-writer	writer.vtu	ParaView output, no VTK linkage
heat-solver, elasticity-stub, modal-stub, cfd-stub	solver.heat.linear, solver.elasticity.linear, solver.modal.linear, solver.cfd.simple	Closed-form demonstration stubs, not FEM/CFD
fenicsx-solver	solver.heat.fenicsx	Opt-in — the one real discretised solve
openfoam-solver	solver.cfd.openfoam.{simple,pimple,inter}	Opt-in, out-of-process
mesh-quality, scalar-magnitude	postproc.mesh_quality, postproc.scalar_magnitude	Post-processing
am-layered-mesher, lattice-reader	mesher.am.layered, reader.lattice	AM geometry
am-thermal	solver.am.thermal.lpbf, postproc.am.melt_pool	Rosenthal (1946) — PHYSICS §1–2
am-distortion	solver.am.distortion.inherent_strain, postproc.am.residual_stress	Keller & Ploshikhin (2014) — PHYSICS §3
am-polymer	solver.am.polymer.fff, postproc.am.bond_strength	FFF interlayer bonding — PHYSICS §4
am-manufacturability	solver.am.{overhang,printability,buildtime}	DfAM checks — PHYSICS §5
am-slicer	writer.am.{gcode,cli,report}	Planar slicing and build reports
marine	solver.marine.{hydrostatic,hull_collapse,corrosion}, writer.marine.qualification_report	Windenburg–Trilling, PREN/CPT — PHYSICS §6–8

Building

Prerequisites: CMake ≥ 3.25, Ninja, a C++20 compiler (GCC 13 / Clang 17 / AppleClang / MSVC 19.36+), and vcpkg cloned with VCPKG_ROOT exported.

On macOS, vcpkg builds libsodium through autotools, so you also need brew install autoconf autoconf-archive automake libtool. Without them the first cmake --preset dev fails inside the vcpkg port build rather than in souxmar's own configure, which makes the cause easy to misread.

cmake --preset dev
cmake --build --preset dev
ctest --preset dev --output-on-failure

Other presets (see CMakePresets.json): dev-python, ci-linux-gcc, ci-linux-clang, ci-macos, ci-windows, asan, tsan.

The first cmake --preset builds vcpkg dependencies from source (~5 minutes for the default feature set, longer with heavy adapters enabled). Later runs use the vcpkg binary cache.

How it's funded (open-core)

The library, plugin SDK, CLI, Python bindings and the desktop app are Apache-2.0 — no crippled community edition. Optional managed services are intended to be commercial; none of them is running today. You bring your own Anthropic / OpenAI / local-Ollama key and your own compute. See docs/BUSINESS_MODEL.md.

Documents

Start here

• docs/PHYSICS.md — every model, its citation, its validity envelope, and its known error
• docs/CAPABILITIES.md — generated capability table
• docs/VISION.md — purpose, target users, scope, non-goals
• docs/ARCHITECTURE.md — layers, data model, dependencies

Reference

• docs/PLUGIN_SDK.md — C ABI, plugin lifecycle, versioning
• docs/AI_INTEGRATION.md — agentic chat, BYOK credentials, tool surface
• docs/MANUFACTURING.md · docs/MARINE.md — capability inputs and model limits
• docs/DESKTOP_APP.md · docs/UI_DESIGN.md — desktop stack and design system
• docs/adr/ · docs/rfcs/ — decision records

Process

• CONTRIBUTING.md · SECURITY.md · docs/GOVERNANCE.md · docs/ROADMAP.md
• docs/HISTORY.md — how the project got here, and an honest note about the shape of its commit history

License

Apache License 2.0. See LICENSE.