radia-mcp
First-and-only public Model Context Protocol (MCP) server suite for Coreform Cubit, Gmsh, build123d, and the Radia CAE ecosystem — including differential geometry and Mathematica integration. Pioneering MCP territory for mesh generators worldwide.
Killer demo (30 seconds): ask Claude to derive the Kelvin transform
factor by hand — it gets stuck on a 3×3 Jacobian + 27-term Laplacian.
Then ask it to use differential-forms + mathematica together:
> verify_with_mathematica(identity="kelvin")
Claude pulls the recipe, sends it to Wolfram, reports back
k=1: Laplacian(psi) = 0 [harmonic — factor R/|y| is correct]
k=2: Laplacian(psi) ≠ 0 [factor wrong]
k=3: Laplacian(psi) ≠ 0 [factor wrong]
— in 8 seconds. The kelvin_factor = R/|y| that takes half a day
in vector calculus appears as the conformal weight λ^((n−2k)/2) of a
k-form, and Mathematica verifies it symbolically.
Authored by the Sugawara Lab (菅原研究室), Kindai University — where the lab-standard primary pair is build123d (CAD authoring) + Cubit (hex meshing), with Gmsh as the post-processing workhorse.
Why this exists
radia-mcp lets an AI assistant drive a real CAE pipeline end-to-end:
build123d script → STEP → Cubit hex mesh → .msh v4.1 / .vol → ngsolve / radia
↑ ↓
AI authors AI iterates safely
(Builder API) (auto-checkpoint
+ batch dry-run
+ live GUI replay)
What makes it different from typical CAD-MCP servers (FreeCAD, OpenSCAD, Blender):
- Test-then-reflect safety pattern: every risky operation
auto-checkpoints to
.cub5, runs in a disposable batch Cubit first, reflects to the live GUI only when batch passes cleanly. The user watches success paths, not failures. - Scheme ladder with geometry split:
cubit_mesh_autowalksauto → sweep → polyhedron → tetmesh, auto-detects compound bodies (volume/surface ratio heuristic) and auto-webcuts before retrying. - Universal mesh backend for the CAD-MCP ecosystem:
any_step_to_cubit_hexaccepts STEP from any upstream MCP (FreeCAD-MCP / OpenSCAD-MCP / cadquery-mcp / Blender-MCP / Onshape-MCP / KiCad-MCP) and dispatches to the same hex-mesh backend. - Deeply scraped knowledge (≈800 lab files + 65 build123d examples
- 15 bd_warehouse + GitHub issues + GitLab issues + StackOverflow + YouTube tutorial transcripts + Coreform training pack), all searchable via tf-idf with heading boost.
- Symbolic verification of FEM formulations (no other MCP server
does this). Pair
differential-forms(theory: Bossavit Whitney complex, Arnold-Falk-Winther FEEC, Kameari edge elements) withmathematica(Wolfram subprocess bridge) to:- Verify d² = 0 on any specific function
- Compute element mass / stiffness matrices on a unit tetrahedron in closed form
- Derive
kelvin_factor = R/|y|for the Kelvin transformation from the conformal-weight formula - Output paper-quality TeX for hand-tuned identities Backed by 21-PDF cohomology bibliography (Bossavit 1998, FEEC 2006, Whitney 1957, Kameari 2011, Codecasa 2010, 新しい計算電磁気学 2003 ...).
Validated EM & multiphysics engineering (radia-ngsolve)
Most CAE-MCP servers stop at generating a script. radia-ngsolve goes further: an
AI assistant can solve a real electromagnetics / multiphysics engineering problem in
open-source NGSolve, and every capability is cross-checked against
a closed-form analytic solution. The result is a library you can trust an AI to drive.
20+ "COMSOL-class" problems, each validated and baked in
Each model ships as a reusable helper + runnable example + regression test + queryable
knowledge (ngsolve_usage(...)), so the server gets smarter,
not just a pile of scripts. Agreement with the closed form is typically well under 1 %:
| Domain | Examples (validation vs closed form) |
|---|---|
| Electrostatics | 3-D capacitance (<0.2 %), layered dielectrics (0.1 %), capacitance matrix (0.03 %), electrostatic force (0.4 %), two-wire line (0.1 %) |
| Magnetostatics | μ-metal shielding sphere/cylinder (0.9–1.3 %), Halbach PM dipole (0.3 %), cylinder magnet on-axis (centre 0.01 %), finite solenoid + Nagaoka inductance (0.5 %), Helmholtz uniformity (<1 %), iron-yoke dipole gap field (0.5 %) |
| Field quality | accelerator-magnet multipoles b_n / a_n — normal-quad main term 0.02 %, allowed/forbidden harmonics resolved |
| Eddy / AC | round-wire skin effect R_ac/R_dc vs Kelvin ber/bei (0.07 %), induction heating |
| Force / torque | weighted Maxwell-stress ("eggshell") force & torque, busbar Lorentz force (~1 %) |
| Multiphysics | electro-thermal Joule heating (exact), electro-thermo-mechanical thermal-stress chain (exact), magneto-mechanical Lorentz→beam deflection vs Euler-Bernoulli (0.02 %) |
Many cases are additionally cross-checked, internally, against a reference commercial FEM solver (three-way agreement analytic = open-source = commercial); the published numbers above are all against the closed-form analytic, the unimpeachable reference.
Designer-facing, not textbook
The helpers take what an engineer actually has — material, geometry, excitation — and
return what they want: capacitance, shielding factor, multipole spectrum, inductance,
force/torque, temperature rise, thermal stress, deflection. Ask the MCP server
ngsolve_usage("field_quality"), ("solenoid"), ("c_magnet"), ("elasticity"), … for
the validated recipe (70+ topics), or lint_radia_script to catch the known FEM traps
before they cost a debug session.
How it's built
Pure-Python NGSolve (H1 / HCurl / axisymmetric H1Henrotte / VectorH1 elasticity), small
reusable solvers chained for couplings, fast pytest contracts for package health, and
explicit validation scripts under validation/ that lock every number above. Open
source (BSD-3); validated against closed-form analytics, with commercial solvers used
only as an internal benchmark.
Demo (placeholder — recordings to be added)
build123d_to_cubit_hex(script=generate_build123d_script("helix_coil")["script"], target_size=1.0)
in one call:
1. build123d Builder API: 4-turn helix + 2 radial leads (3 prismatic bodies)
2. STEP export → /tmp/coil.step
3. cubit_batch_try (headless): scheme ladder
├─ auto → 1668 hex / 0 tet ✓ WIN
└─ (sweep / polyhedron / tetmesh skipped)
4. live Cubit GUI replay → user sees the winning recipe execute
5. .cub5 checkpoint preserved for rollback
Total time: ~30 s on a 4-core laptop. AI watches state delta, applies hint heuristics on failure, and never leaves a half-broken mesh in the user's window.
Install
pip install radia-mcp # core (Cubit + build123d + GMSH lint/visualization + interop)
pip install radia-mcp[build123d] # adds build123d itself
pip install radia-mcp[cadquery] # adds CadQuery (interop with cadquery-mcp)
pip install radia-mcp[gmsh] # adds gmsh Python bindings for optional GMSH workflows
pip install radia-mcp[youtube] # adds youtube-transcript-api (tutorial scrape)
pip install radia-mcp[radia] # adds Radia core (radia-coupled servers)
pip install radia-mcp[full] # everything above
Install the standalone MATLAB distribution and the official Optuna MCP server
separately; neither is a runtime dependency of radia-mcp:
pip install --upgrade radia-optuna
pip install --upgrade optuna optuna-mcp
The ownership rule is strict: the official optuna/optuna-mcp live
tools/list owns every shared Study/Trial/query/visualization/Dashboard
operation it exposes. mcp-server-radia-matlab supports only the MATLAB
differences through matlab_optuna_mcp_route: table/MAT persistence, Simulink
monitoring and failure telemetry, MATLAB parallel execution, the required
20-command optuna_mex, and Radia CAE artifact adapters. Seeded numerical
oracle checks execute pinned optuna==4.9.0 directly because the verified
upstream MCP sampler tool does not expose a seed.
radia-optuna is independent and unofficial; it is not affiliated with,
sponsored by, or endorsed by Preferred Networks, Inc. or the Optuna project.
Optuna, the Optuna logo and any related marks are trademarks of Preferred Networks, Inc.
Radia does not use the Optuna logo or present its MCP layer as
official. Optuna and optuna-mcp are MIT-licensed upstream projects; their
copyright/license notices are bundled in the radia-optuna wheel's
THIRD_PARTY_NOTICES.md. Oracle regeneration starts the official MCP locally
over stdio with a fresh temporary SQLite database. Routine tests use the checked
fixture, do not launch Dashboard, do not touch shared/production storage, and
do not automatically open upstream issues or pull requests.
Requires Python ≥ 3.10. Coreform Cubit is auto-discovered from
standard install paths; set CUBIT_BIN_DIR env var if installed
non-standard.
Release Quality Stance
radia-mcp treats a green test matrix as release-candidate evidence,
not as the final operational claim. The public MCP surface is healthy
when the matrix, policy lint, version consistency, generated
docs/TOOLS.md drift check, and top-level pytest collection all pass.
Operational quality is claimed only after PyPI-installed MCP entry
points smoke successfully and the release-quad machine checks pass on
the lab deployment roles. Public-safe quality records live in
validation/mcp_quality/.
★ Discovery — start here
Before calling any specific server, ask mcp-server-radia-meta which
server has the knowledge you need. It is the authoritative cross-server
catalog of all 49 radia-mcp servers and answers "which tool covers
concept X?" without trial-and-error.
# catalog, health, golden-gate, and bug-pattern tools
radia_mcp_overview() # all 49 servers + live tags
radia_mcp_get("bayesian-opt") # full info for one server
radia_mcp_by_tag("optimization") # filter optimization/theory servers
radia_mcp_related("bayesian-opt") # cross-link map for optimization servers
radia_mcp_health() # importability probe of all 49
radia_mcp_golden_gate() # catalog/discovery/public-boundary gate
Then drill into a specific server with its <short>_status() (auto-
introspected tool list + dep probe) and <short>_topics() (for
dispatcher-style servers: the topic enum) tools.
Pattern: 3-call discovery instead of guess-and-error —
radia_mcp_by_tag("optimization")
→ [bayesian-opt, evolutionary, topology-optimization,
data-assimilation, gnn, pinn]
→ bayesian_opt_status() # confirm radia-side theory tools
→ bayesian_opt_topics() # topic enum for BO / GP / FMQA
→ official optuna-mcp # every shared operation in live tools/list
→ matlab_optuna_mcp_route() # MATLAB/Simulink differences only
MCP servers
The catalog is the source of truth — call radia_mcp_overview() for
the full 49-server live list. The historically primary servers are
shown below for reference; everything else is discoverable via meta.
Standalone (no Radia core dependency — pip install radia-mcp)
| Server | Entry point | Tools | Highlights |
|---|---|---|---|
| ★ meta | mcp-server-radia-meta |
9 | Cross-server catalog + health/golden gate — RECOMMENDED FIRST CALL |
| literature-index | mcp-server-literature-index |
9 | Full-text search across 2,339 lab literature files in W:/03_文献・論文 (ChromaDB + semantic search) |
| Cubit | mcp-server-cubit |
45 | cubit_mesh_auto, cubit_exec_safely, cubit_ask, scheme ladder + geometry split, .cub5 checkpoint/restore, scrape index over Coreform forum + S:\CoreformCubit lab archive (787 files) + YouTube + Coreform training |
| build123d | mcp-server-build123d |
29 | build123d_to_cubit_hex, lint_build123d_script, build123d_try (subprocess isolation), build123d_inspect_step, build123d_heal, build123d_api, Radia/general templates, CadQuery + bd_warehouse interop |
| GMSH | mcp-server-gmsh |
10 | lint_gmsh_script, gmsh_audit_summary, gmsh_numsubedges_remediation_plan, gmsh_mesh_generation_remediation_plan, references + examples |
| Force | mcp-server-force |
21 | Common Motor/MagLev force layer: shared result normalization; static and peak/RMS phasor Lorentz/Maxwell force and torque; virtual work, coenergy and uniform/sampled air-gap torque; method selection, independent-method/action-reaction/lift-weight gates; and validation guidance (numerical tools require the radia extra). |
| differential-forms | mcp-server-differential-forms |
15 | Visual differential geometry for computational EM: intrinsic metric, curvature/holonomy, Cartan moving frames, k-forms, exterior derivative, Hodge star, Whitney complex, de Rham, tree-cotree, FEEC, and executable geometry/gauge gates. Distilled from Needham 2021/2026, Bossavit 1998, Whitney 1957, Kameari 2011, Arnold-Falk-Winther 2006, 新しい計算電磁気学 2003, and Codecasa 2010. |
| mathematica | mcp-server-mathematica |
13 | Wolfram Mathematica subprocess bridge: evaluate expressions, execute tracked verification scripts with JSON reports, batch named identities in one kernel, and select a course/differential-forms/paper verification workflow, plus simplify, TeX, vector calculus, units, solve, integrate, differentiate, and status helpers. Pairs with differential-forms for symbolic verification of d²=0, Stokes, Whitney elements, Kelvin transform, Maxwell identities. Requires wolframscript on PATH. |
Radia-coupled (pip install radia-mcp[radia])
| Server | Entry point | Highlights |
|---|---|---|
| radia-ngsolve | mcp-server-radia-ngsolve |
NGSolve FEM/BEM, Whitney elements via H1/HCurl/HDiv, PEEC inductance, closed-form formulas (Wakao-Igarashi Part 1-9, cuboid average B, Bessel impedance, etc.) |
| ih | mcp-server-ih |
Induction heating workflow: workpiece SIBC, ESIM nonlinear cell problem, Karl iteration, screening physics |
| peec | mcp-server-peec |
PEEC: Loop-Star, FastHenry, PyPEECBuilder, Bessel/Dowell/ESIM SIBC, PRIMA model-order reduction, SPICE extraction |
| electromagnet | mcp-server-electromagnet |
Accelerator magnets: magnetic rigidity and beam-optics handoff, normal/ramped/superconducting design, field measurement, curated textbook guide, CoilBuilder, Hantila, hysteresis, IMA, and multipoles |
Claude Code / Desktop configuration
Add to your MCP client config (Claude Desktop, Claude Code, Cursor, Continue, …):
{
"mcpServers": {
"radia-meta": {"command": "mcp-server-radia-meta"},
"literature-index": {"command": "mcp-server-literature-index"},
"cubit": {"command": "mcp-server-cubit"},
"build123d": {"command": "mcp-server-build123d"},
"gmsh": {"command": "mcp-server-gmsh"},
"force": {"command": "mcp-server-force"},
"differential-forms": {"command": "mcp-server-differential-forms"},
"mathematica": {"command": "mcp-server-mathematica"},
"radia-matlab": {"command": "mcp-server-radia-matlab"},
"optuna": {
"command": "optuna-mcp",
"args": ["--storage", "sqlite:///C:/temp/optuna_mcp.db"]
}
}
}
The path above is an example of user-owned persistent local storage. Automated oracle tests must instead create a unique database in the per-run temporary directory and delete it after the local stdio server exits.
Registering more than the 9 above is rarely necessary — once you have
meta, radia_mcp_get(name) returns the entry point for any of the
49 servers and you can register them on demand. The full list of
catalog-driven server names is in
radia_mcp.meta.catalog.CATALOG.
For local development from a checkout (no install needed):
{
"mcpServers": {
"cubit": {
"command": "python",
"args": ["-m", "radia_mcp.cubit.server"]
}
}
}
Knowledge bases shipped (offline)
| Server | Bundled knowledge | Lines |
|---|---|---|
| Cubit | cubit_api_reference.py (600+ functions) + scripting + forum tips + netgen workflow + export rules + panels |
≈ 29 000 |
| build123d | Auto-generated API reference (inspect.getmembers, 142 classes / 65 functions) + 18 curated topics (Plane/Axis/Location cookbook, Builder ↔ Algebra rosetta, joints, assemblies, CAE workflow, …) |
1 673 |
| GMSH | Visualization/post-processing policy, MSH v4.1 spec, high-order display guidance, lint rules, examples, and remediation planners | 2 008 |
Plus persistent failure log per kind, fed into every *_lookup /
*_ask retrieval so past mistakes are searchable next session.
Live-scraped knowledge (cached 7 days)
| Source | Kind | Volume |
|---|---|---|
| Coreform forum (Discourse) | cubit_examples |
full archive walk, ≈ 60 code-bearing posts |
S:\CoreformCubit + Radia/01_GitHub/examples |
cubit_examples |
787 .jou / .py files (lab archive + Coreform training pack) |
| Cubit YouTube tutorials | cubit_examples |
5 transcripts |
GitHub .jou code search (PAT) |
cubit_examples |
15 files |
gumyr/build123d/examples |
build123d_examples |
65 scripts |
gumyr/bd_warehouse |
build123d_examples |
15 modules |
gumyr/build123d Issues + GraphQL Discussions |
build123d_examples |
60 + 50 |
| build123d YouTube | build123d_examples |
4 transcripts |
gitlab.onelab.info/gmsh/gmsh issues |
gmsh_examples |
walking 3000+ |
StackOverflow / SciComp.SE [gmsh] |
gmsh_examples |
30 votes-sorted |
| gmsh YouTube | gmsh_examples |
5 transcripts |
{cubit,build123d,gmsh}_examples_refresh() re-scrapes anytime;
otherwise cache TTL is 7 days.
Lab stance (Sugawara Lab, 菅原研究室)
| Tool | Position |
|---|---|
| build123d | 主力 (push) — new lab work is authored here |
| Cubit | 主力 (push) — hex mesh + visualization |
| FreeCAD | 応援 (friendly) — first-class interop, not first-class authoring; lab respects the FreeCAD community |
| CadQuery | interop / compat — OCCT sibling |
| OpenSCAD | interop / compat — legacy scripts |
The MCP-tool layer encodes this stance: list_cad_mcp_interop reports
primary_pair: "build123d (CAD authoring) + Cubit (hex mesh)" and
flags FreeCAD as friendly, others as compat.
Design pattern: Checkpoint-Batch-Commit-Race (CBCR)
The signature workflow cubit_mesh_race_with_human realizes a
publishable design pattern we call CBCR: the user and N AI
agents race the same starting state; first-to-finish wins; the
human's in-progress work is never silently overwritten.
Full design document (defensive prior-art publication, BSD-3-Clause):
docs/design/checkpoint_batch_commit_race.md.
Architecture (Plan A)
The Cubit server uses Plan A: launch
coreform_cubit.exe -nojournal cubit_bootstrap.py, which installs a
PySide6 QTimer (200 ms poll) inside Cubit's Qt event loop. The MCP
side drops *.req.json into a temp directory; the bootstrap polls,
runs cubit.cmd() on the Qt main thread, and writes out/*.resp.json
back. No sockets, no pipes — just atomically-renamed JSON files.
This sidesteps every ABI / event-loop / GIL issue the COMSOL-style "external Python + library load" approach hits.
cubit_exec_safely adds a second safety layer on top:
- Auto-
save asthe live GUI state to~/.cubit_viewer/checkpoints/autosafe_<ts>.cub5. - Spawn a fresh batch Cubit,
openthe checkpoint, run candidate commands. - If batch passes (and
cubit.get_error_count()didn't tick up — silent-error guard), replay on the live GUI. - Otherwise: live GUI untouched, checkpoint label returned for
cubit_restore.
Quickstart
build123d → Cubit hex mesh, one call
build123d_to_cubit_hex(
script=generate_build123d_script("helix_coil")["script"],
target_size=1.0,
prefer="hex",
commit_to_gui=True,
)
# ⇒ STEP exported → batch ladder picks scheme auto → live Cubit GUI
# shows 1668 hex / 0 tet / 3780 nodes
Safe Cubit live edit
cubit_exec_safely(commands=[
"volume all size 1.0",
"volume all scheme tetmesh",
"mesh volume all",
])
# ⇒ auto-checkpoint → batch dry-run → only on pass: live GUI mesh
# if dry-run fails, GUI is untouched + rollback label returned
Search any layer of knowledge
cubit_ask("hex meshing tutorial")
# ⇒ unioned ranked hits across bundled KB + lab archive (787 files)
# + Coreform forum + YouTube transcripts
Audit GMSH visualization policy
gmsh_audit_summary("examples") # machine-readable policy audit
gmsh_numsubedges_remediation_plan("examples") # high-order display companions
gmsh_mesh_generation_remediation_plan("examples")
gmsh_reference("all") # MSH / API / display reference
Contributing
Bug reports + PRs welcome — particularly for:
- Additional scrape sub-sources under
radia_mcp.common.examples(mailing list archives, more YouTube channels, blog posts). - Cookbook topics for
build123d_usage/gmsh_usage/gmsh_reference— worked-example knowledge fragments are always welcome.
See CONTRIBUTING.md for the lightweight workflow. History of every release is in CHANGELOG.md.
License
BSD-3-Clause — see LICENSE.
Acknowledgments
- Coreform for Cubit and the open Discourse forum that we scrape with appreciation.
- Roger Maitland (gumyr) for build123d and bd_warehouse — the lab-standard upstream CAD.
- Christophe Geuzaine for Gmsh and the open issue tracker on
gitlab.onelab.info. - MCP / Anthropic for the protocol that made AI-driven CAE practical.
1 maintainer
Release files for radia-mcp 1.4.46
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| radia_mcp-1.4.46-py3-none-any.whl | Python 3 | none | any | Details |
Release files / radia_mcp-1.4.46-py3-none-any.whl
| Download URL | radia_mcp-1.4.46-py3-none-any.whl |
|---|---|
| Size | 6.1 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
3379ac05a303b87960c64072b9fce87b3a534f49ea7ee401fa175ec6b0e96636
|
|
BLAKE2b-256 checksum How to use checksums |
450c31d46a0c34b365c9ab0f415541687418f852e28866eec134120a5f868289
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Aug 27, 2026.
Transparency log
