HOW INTACT WORKS

A conversation.
A complete engineering pipeline.

Explore the local CAD architecture, open-source engines, modeling flow and evidence behind every accepted revision.

How Intact comes to life

Intact Simulation connects a conversation in ChatGPT to a local engineering workspace. ChatGPT interprets the design brief and chooses tools. Intact coordinates those tools, builds real geometry, records revisions and returns measured evidence. The CAD kernel constructs the shape; a numerical solver computes supported analysis results. The viewer makes both inspectable.

The system brings together Python engineering tools, a typed Model Context Protocol (MCP) interface, isolated workers, versioned project storage and a browser viewer. Intact’s contribution is the workflow between them: preserving design intent, binding edits to actual geometry, checking exports and keeping results associated with the revision that produced them.

System architecture · AI reasoning, local execution and visible evidence
Your brief in ChatGPTPurpose, dimensions, references, interfaces and acceptance criteria
Intact plugin · MCP tool interfaceValidated inputs · current model identity · revision guards · workflow instructions
Native CAD workerPython → build123d → OpenCASCADE
Structured feature workerFreeCAD constraints and explicit feature graph
Mesh & numerical workersTPMS sampling · meshing · supported PDE solves
Audit, requirements and revision-bound artifactsCheck geometry and exports; reject known failures; retain unverified criteria
Local viewer, reports and downloadsInspect → measure → select → Ask AI → guarded next revision

Three boundaries that matter

Conversation
Your signed-in ChatGPT client provides AI reasoning. Intact does not contain trained language-model weights and does not retrain ChatGPT on your models.
Local engineering
CAD and supported numerical work execute in the installed native Windows runtime. Linux source installations retain their own isolated runtime. Projects and runtime files are separate from the installation package.
Public website
The Finished Work viewer displays prebuilt examples and recorded evidence. New custom CAD and new numerical studies run through the installed plugin.

The flow from a geometry request to a saved model

The workflow applies to supported parts, assemblies and sampled structures. “Any geometry” is a design ambition, not a guarantee that every requested shape or operation can be constructed. Essential missing dimensions or ambiguous references must be resolved before committing a design.

  1. Define the brief. State units, the envelope, components, hole patterns, mating interfaces and intended output. Separate exact dimensions from assumptions.
  2. Save checkable design requirements. Record supported checks for dimensions, placement, component count, holes and clearances. Missing criteria remain unverified.
  3. Choose a representation. Use native solids for exact CAD operations, constrained sketches and a feature graph for explicit parametric relationships, or a sampled mesh for implicit TPMS geometry.
  4. Prepare the build. ChatGPT supplies a bounded Python program and named numeric parameters, or structured sketch/feature operations. Existing designs retain their model identity, unrelated parts and revision guard.
  5. Execute in an isolated worker. The server checks input policy and dispatches computation. Generated CAD Python runs in Windows AppContainer with bounded Job Objects, or Linux bubblewrap for source installations, with filesystem/network restrictions, resource limits and a timeout.
  6. Audit the candidate and exports. Check topology, finite coordinates, positive volumes, mesh integrity where applicable and preservation through export. Only bounded repairs that preserve geometric invariants are allowed.
  7. Evaluate saved requirements. Measure supported acceptance criteria against native exported geometry. A known failed requirement blocks publication of that proposal.
  8. Decide: publish or repair. A successful proposal becomes a new revision. A rejected build preserves the last saved design; repair the failed feature without removing requirements merely to pass.
  9. Inspect and deliver. Compare measured facts and rendered views with the brief. Export the appropriate native or mesh formats and disclose assumptions and unverified checks.
The repair loop: a failed check returns to the relevant build or input step. It does not turn a red result green by changing a tolerance, removing a feature or publishing a different shape.

The engineering stack and each tool’s job

Core components in Intact Simulation 0.19.0
LayerTechnologyWhy it is used
Tool contractsPython, MCP, PydanticTyped requests, explicit operations, validated settings and structured responses.
Native geometrybuild123d 0.13, OpenCASCADESolids, booleans, holes, fillets, assemblies and native STEP geometry.
Constraints & historyPinned FreeCAD runtimeActual sketch-solver diagnostics and explicit supported feature rebuilds.
IsolationWindows AppContainer + Job Objects; Linux namespaces/bubblewrapA restricted filesystem and network boundary for generated Python; AST policy checks add another layer.
Numerical arraysNumPy, SciPySampling, matrix operations, sparse solves and numerical diagnostics.
Implicit geometryscikit-image, trimesh, Manifold3DSampled isosurfaces, mesh inspection and bounded mesh operations.
Finite elementsGmsh, scikit-fem, meshioTetrahedral meshing, supported finite-element formulations and result exchange.
Local serviceFastAPI, UvicornLoopback viewer delivery, model requests and job/result transport.
Interactive displayJavaScript, Three.jsOrbit, sectioning, native-reference selection, component controls and result fields.
Reports & exchangepython-pptx, ReportLab, ezdxf, LibreDWG, OpenUSDReview decks, documents and supported drawing/scene formats.
Repeatable installationuv, locked dependencies, checksum-pinned downloadsConsistent environments and explicit runtime versions.
Regression evidencepytest, JavaScript tests, Playwright ChromiumBackend checks, interaction contracts, browser behavior and export inspection.

Optional PicoGK adds a pinned voxel-based TPMS path through its Linux runtime and .NET. Its output is a mesh approximation with voxel-resolution evidence. It does not create an exact smooth STEP surface.

Commands ChatGPT uses on your behalf

Users can describe the outcome in ordinary language. The assistant selects the relevant tool; there is no need to paste backend commands into the chat. The installed catalog exposes 65 MCP tools in version 0.19.0.

TaskRepresentative tool calls
Check readiness and recover a projectstatus, list_models, get_model
Create or change CADbuild_model, update_parameters, import_cad
Preserve design intentsolve_sketch, edit_feature_tree, set_design_requirements
Measure and checkmeasure_model, validate_model, verify_design, audit_model
Try an alternativepreview_parameters, parameter_configuration, apply_configuration
Inspect and sharesnapshot_model, export_model, create_presentation
Run a supported studyconfigure_simulation, submit_engineering_job, run_virtual_analysis

How the Auditor and V&V improve confidence

Accuracy begins with the right requirements and a valid representation. Auditing detects invalid geometry and broken exports. Design verification checks supported requirements against measured geometry. Validation asks whether the geometry, assumptions and study represent the real intended use. These are related, but passing one does not establish the others.

1. Check the geometric object

For native CAD, the audit examines B-rep validity, finite positive volumes, dimensions and tessellation. For sampled meshes, it checks finite coordinates, watertightness, winding and positive volume. Bounded kernel healing or orientation corrections must preserve counts, bounds, volume and area within strict invariants; larger changes require revising the model.

2. Check the files you actually receive

Export checks reopen STEP geometry and inspect preservation of part identity, dimensions and volume. GLB structure and 3MF integrity are checked too. A beautiful viewport cannot compensate for an export that drops a part or changes the envelope.

3. Check against design intent

Saved requirements can test supported bounds, placements, counts, hole patterns, clearances and sampled wall-thickness locations. Native geometry supplies the evidence. “Six holes” must mean six holes at the requested centers and diameters, not merely six items with convincing names. Finite thickness samples do not certify a global minimum wall thickness.

4. Keep status honest

Passed

The stated supported check met its criterion.

Failed

The measured result violated a criterion; a known failed design requirement blocks that proposal.

Unverified

The criterion is missing, unsupported, incompletely sampled or lacks sufficient evidence.

Review opens the measured envelope, component count and audit context. Its AI checks and alternative ideas are read-only requests. Suggestions become geometry only through a subsequent requested edit. Human visual review and suitable external engineering validation remain necessary.

Simulation has its own evidence chain

Mechanical Analysis supports bounded 3D steady conduction, small-strain linear elasticity and creeping-flow Stokes problems. A study needs the actual solid, material, selected boundaries, units and physical conditions. Geometry is audited before solving; result audits inspect finite fields, residuals, conservation, refinement and supported regimes before a report is published.

An unrefined solution remains unverified. A failed numerical check is not a successful solve. Display deformation is a visualization scale; it does not change the CAD. Virtual Analysis provides kinematic motion previews, not gravity, collision, tire traction or suspension-contact dynamics.

Reliable changes leave a trail

Stable part names, guarded revisions and artifact-bound face/edge references reduce accidental edits to the wrong geometry. Results carry their source identity. History previews are read-only; restoring is an explicit action. Parameter previews rebuild in scratch space, so exploring an idea need not overwrite the accepted model.

Intact does not claim universal CAD generation, whole-assembly contact, turbulent CFD, nonlinear mechanics or certification. Open-source engines make the computation inspectable; requirements, regression tests and measured evidence make its limits visible.

Explore further

Follow the ChatGPT tutorial, read the engineering guides, or set up the plugin.

Engine references: build123d · OpenCASCADE · FreeCAD · Gmsh · scikit-fem · Three.js. Dependencies retain their respective licenses.

Install once.
Design through conversation.

Download Intact v0.20.0 for Windows. Install privately in Codex with your own account; CAD runs on your computer.

Follow the installation procedure →
Download Windows ZIP ↓