Kratos-MCP-Server：让 AI 助手端到端驱动 Kratos 有限元仿真

An MCP server that lets AI assistants drive Kratos Multiphysics finite element simulations end to end:

Introspect the installation: applications, elements, conditions, constitutive laws, variables, solvers and processes and their default parameters (process defaults are parsed from the Kratos source, no build needed).
Scaffold simulation cases from templates: structural (static/dynamic/modal), thermal (transient/stationary), fluid (monolithic or fractional-step incompressible) and potential flow — plus multi-stage orchestrated cases that chain analyses, CoSimulation cases where several solvers exchange interface data inside each step (weak/strong Gauss–Seidel or Jacobi coupling) and partitioned FSI cases (fluid + structure + ALE mesh motion across a mapped interface, Dirichlet–Neumann iteration). Curated material and linear-solver presets included. ProjectParameters.json, Materials.json and structured MDPA meshes with named boundary regions.
Build and edit meshes and cases: add, move and delete nodes, elements, conditions and geometries; create sub model parts from geometric selectors; transform, merge, extrude, convert quads/hexes to simplices, extract boundary conditions on free faces and check mesh quality; parametric shapes (annulus, disc, cylinder, sphere, polygon), meshio import/export and, with the optional gmsh extra, arbitrary geometry. Edit ProjectParameters and Materials generically: get/set any setting, add/update/remove any process, stage-aware (mesh, geometry, project).
Remesh adaptively with MMG (MeshingApplication): mesh_remesh refines or coarsens an .mdpa around a region, a level set, an analytic field or a solved VTK result field (hessian metric, transferred field written as NodalData), keeps the named boundaries and reports the field transfer error; add_remeshing_process adds a setup-time mmg_process to a case (tools, tutorial).
Study parameter sweeps and structured mesh convergence with isolated variants, a bounded background queue, and physical-location probe responses (tools, tutorial).
Run simulations (single- or multi-stage) as managed background jobs (status, live logs, progress, cancel) that survive server restarts, with optional immutable input snapshots, reproducible reruns, checkpoint resume for isolated serial single-stage jobs, and local MPI launch (mpi_ranks, omp_threads) with per-rank logs and whole-group cancellation.
Drive interactively: open a live session (a Kratos Model + AnalysisStage kept in memory), advance it a few steps at a time, inspect fields, nodes and elements between steps, pin or change nodal values, and continue — the direct, stateful route into Kratos' Python layer next to the batch-job route. Optional raw Python access behind KRATOS_MCP_ALLOW_EXEC=1 (tools, tutorial).
Post-process VTK results: summaries, point probes, convergence analysis, physical-time histories, CSV export, and reference comparisons.
Preview results without ParaView: PNG screenshots and GIF animations (deformed shapes, field contours) rendered with pyvista and shown inline in the conversation — optional viz extra.
Interoperate: explain an existing ProjectParameters.json, and convert cases to/from the Kratos FlowGraph visual node editor (lossless round-trip).

105 tools, 21 resources (15 worked examples) and 8 guided prompts. See the full documentation in docs/ (VitePress).

Quick start

A local Kratos build is not required — the server can pip-install Kratos itself on first use.

Once published to PyPI, no clone needed — uvx fetches and runs it:

claude mcp add kratos -- uvx kratos-mcp-server

From a local checkout (current state, before the first PyPI release):

git clone https://github.com/loumalouomega/Kratos-MCP-Server
cd Kratos-MCP-Server
uv sync
claude mcp add kratos -- uv --directory "$PWD" run kratos-mcp

Then ask your assistant something like:

Check the Kratos installation — install it if it's missing — then set up a cantilever plate 1 m × 0.2 m fixed on the left with a 1 MN/m downward load on the right edge, run it, and report the tip deflection.

The assistant calls kratos_install the first time and reuses it afterwards. If you already have a compiled Kratos checkout, skip that and point KRATOS_ROOT at it instead (-e KRATOS_ROOT=/path/to/Kratos on the claude mcp add line) — see Installation.

Notebooks

Seven notebooks drive the server interactively as an MCP client — no AI assistant involved — each touching most of the relevant tools, resources and prompts in one sitting. Run any with uv sync --extra viz --group dev (adds ipykernel + pyvista) and open it in Jupyter/VS Code against that .venv.

notebooks/cantilever.ipynb: the structural cantilever case — installation introspection, mesh generation, scaffolding, a background job you poll while it runs, VTK post-processing, a rendered PNG and an animated GIF, and cancelling a job in flight.
notebooks/naca_airfoil.ipynb: a NACA0012 airfoil in incompressible laminar flow — reuses a real ~21k-node airfoil mesh from Kratos's own examples repo (simplified physics; see the notebook/tutorial for what and why), computes lift/drag by summing REACTION over the airfoil surface, and renders/animates the pressure field cropped to the airfoil with the newer crop_bounds option.
notebooks/fluid_cavity.ipynb: the lid-driven cavity CFD benchmark — introspect the fluid solver, run the shipped example, probe the velocity field, and render an inline PNG and GIF of the recirculating vortex.
notebooks/materials.ipynb: the materials surface — material & linear-solver presets, process-defaults introspection, and a single-element von Mises plasticity run showing the elastic→plastic transition.
notebooks/multistage.ipynb: multi-stage orchestration — create_multistage_project chaining two load steps, explain_project_parameters, and a lossless round-trip to/from a Kratos FlowGraph node graph.
notebooks/interactive_session.ipynb: interactive sessions — step a load-ramped cantilever two steps at a time, inspect fields and nodes between steps, pin a DOF mid-run, peek into the live Kratos objects, and verify the stepped result equals a batch run.
notebooks/interactive_fluid.ipynb: stepping the lid-driven cavity in chunks and watching the flow approach steady state through live field summaries.
docs/tutorials/potential-flow.md: a runnable NACA0012 perturbation potential-flow benchmark with upstream reference values.

Requirements

Python ≥ 3.10, uv
Kratos Multiphysics, either pip-installed via kratos_install (Linux/Windows x86_64 only — no macOS wheels) or a compiled build (tested with 10.4, StructuralMechanics / ConvectionDiffusion / FluidDynamics / LinearSolvers applications)
Optional, for results_render/results_animate: the viz extra (uv sync --extra viz or pip install 'kratos-mcp-server[viz]') and a working OpenGL context (on headless machines: Xvfb or OSMesa VTK wheels — see docs/tools/visualization.md)

Architecture in one paragraph

Kratos is never imported in the server process — it prints a banner on import (which would corrupt the stdio JSON-RPC stream) and can abort the process on solver errors. Short operations run in a worker subprocess that returns JSON through a result file; simulations run detached with per-job directories under ~/.kratos-mcp/jobs/; interactive sessions keep a live AnalysisStage in a long-lived worker driven over JSON lines. See docs/guide/architecture.md.

Development

See the roadmap for planned work and the changelog for release history.

uv run pytest -m "not kratos"   # unit tests (no Kratos needed)
uv run pytest -m kratos          # integration tests against the real build
npm install && npm run docs:dev  # documentation site

License

MIT