CAD Agent

CAD Agent helps coding agents turn mechanical requirements into validated FreeCAD models. It combines requirement reasoning, knowledge reuse, standard-part provenance, deterministic CAD state, automatic validation, correction, final confirmation, and reusable Design Lessons.

Nothing here guesses a number. Sizing follows published standards, the result carries its own assumptions and limitations, and a finished model is confirmed by a process the agent cannot reach or forge. A learned surrogate may offer a fast screening estimate, but only as a calibrated interval with declared coverage that is refused outside its fitted domain and can never gate completion.

The package provides the mech-cad-design CLI and mech-cad-design-mcp server. A compatible coding agent performs the design reasoning, while an external FreeCAD GUI MCP performs interactive CAD work. The package does not embed a language model and does not replace engineering review.

Design process

User request
  → requirement clarification
  → short design proposal
  → one natural-language direction approval
  → knowledge retrieval
  → CAD modeling
  → automatic validation and correction
  → correction capture
  → final result
  → natural-language final confirmation
  → automatic Design Lesson evaluation
  → finish, or one decision before durable lesson publication

Related MCP server: IntentForge MCP Server

Core capabilities

• Create new designs or edit read-only snapshots of existing FCStd/STEP models.

• Retrieve matching Product Family Knowledge and Design Lessons when available.

• Continue CAD work when knowledge has no match or its backend is unavailable.

• Model interactively in FreeCAD and keep FCStd as the source of truth.

• Size a spur gear drive from duty inputs (power, speeds, material, duty, life, safety factor) through ratio, module, forces, bending and contact stress, shaft diameter, and required bearing capacity, then model and validate the resulting pair against the same calculated values. Preliminary sizing evidence, not a strength certification; see its limitations.

• Find purchasable standard parts through configured structured providers and, when they miss, extend the search to authoritative manufacturer, standards body, industry association, and attributable authorized-distributor sources.

• Register selected CAD components with provider, manufacturer, part identity, source, license, validation evidence, and SHA-256 provenance.

• Calculate from Shigley's Mechanical Engineering Design independently of any one component: stress states and principal stresses, static and fatigue failure criteria, deflection and columns, shafts, bolted joints, springs, rolling and journal bearings, gears, welds, clutches, brakes and belts. Pure standard library, with every shipped constant checked against an independent anchor and every curve fit marked as one.

• Record a surrogate screening estimate produced outside this package, as a calibrated interval with its coverage, calibration provenance and declared training domain. An estimate whose query falls outside that domain is refused and the refusal is kept. Screening never gates completion and never counts as a passed validation check; see Surrogate screening.

• Test a screening interval against the closed-form mechanics result for the same load case, so a miscalibrated surrogate is caught by the analytical answer rather than trusted over it.

• Validate geometry, dimensions, placements, interfaces, assemblies, fasteners, BOM consistency, and visual evidence.

• Bind completion to the exact FCStd SHA-256 and passed JSON, Markdown, and PNG evidence.

• Re-verify the finished model in a separate process the agent does not control, under a SHA-256-pinned FreeCAD executable running digest-pinned scripts, and accept the result only when a host-issued nonce and the recorded model digest both come back unchanged. The agent writes its own validation report; this is the part of the evidence it cannot author.

• Record every validation attempt in an append-only correction ledger, so a failure that was fixed is not lost when the next attempt is recorded.

• Evaluate reusable lessons automatically after the user confirms the final model, including lessons derived from the mandatory checks this design failed and then corrected.

• Feed published correction lessons back through ordinary knowledge retrieval, so a later design finds the defect before repeating it.

• Store long-term Product Family profiles, Knowledge Assertions, and Design Lessons in a local SQLite database by default, with optional PostgreSQL for shared team use and an optional rebuild-only Neo4j projection.

MCP surfaces

The default design surface contains the complete design flow:

• design_system_status

• design_start

• design_status

• design_knowledge_retrieve

• design_record_result

• design_mistakes

• design_gear_size

• design_screening_record

• design_screening_status

• design_confirm

• design_lesson_decide

• standard_part_providers_get

• standard_part_sources_status

• standard_part_download_register

The separate knowledge-admin surface manages Product Family onboarding, knowledge search, Design Lesson supersession or revocation, and explicit Neo4j projection rebuilds.

Architecture

Design sessions live under designs/<design-id>/ as atomic JSON state, one authoritative model.FCStd, optional source snapshots, validation evidence, outputs, and an optional lesson review card. CAD creation and validation do not depend on PostgreSQL.

The knowledge store holds only durable Product Families, Knowledge Assertions, and Design Lessons, in local SQLite by default or PostgreSQL when configured. Neo4j is optional, rebuildable, and never authoritative. See Architecture and trust boundaries.

Learning from corrected mistakes

Every call to design_record_result appends one entry to the design's append-only correction ledger: the model hash, the validation outcome, and each failed check. Nothing in a later attempt rewrites an earlier one, so the record of what went wrong survives the fix.

When the user confirms the final model, the package groups the mandatory checks that failed on earlier attempts and passed on the confirmed model. Each such defect becomes one deterministic Design Lesson candidate carrying origin: validation_correction, its check signature, and how many attempts it cost. These candidates join any the agent proposes on the same immutable review card and follow the same single publication decision.

Derivation runs without a language model and never blocks: a design that made no mistakes derives nothing, an advisory-only failure derives nothing, and a malformed derivation is dropped rather than holding up a completed model. Once published, correction lessons are ordinary Design Lessons, so design_knowledge_retrieve returns them to later designs in the same scope.

Install and run

Python 3.12 or newer is required. There is no PyPI release yet, so install from a clone:

git clone https://github.com/bloodreaper005/cad-agent
cd cad-agent
python -m pip install .

mech-cad-design init \
  --workspace /path/to/mech-cad-design-workspace \
  --actor engineer \
  --organization example-org \
  --design-group example-group
mech-cad-design knowledge bootstrap \
  --workspace /path/to/mech-cad-design-workspace
export MECH_DESIGN_WORKSPACE=/path/to/mech-cad-design-workspace
mech-cad-design-mcp

The knowledge store is a local SQLite database inside the workspace, so no service has to be running. mech-cad-design status reports workspace, FreeCAD, and knowledge readiness as structured JSON and exits non-zero when setup is incomplete.

Windows PowerShell:

mech-cad-design init --workspace "D:\Mechanical Design Workspace" --actor engineer --organization example-org --design-group example-group
$env:MECH_DESIGN_WORKSPACE = "D:\Mechanical Design Workspace"
mech-cad-design-mcp

Command line

Every command prints one JSON document and exits 0 ready, 1 warning, 2 setup required, or 3 blocked. gear size uses the same scale: 0 sized, 1 sized with warnings, 3 rejected.

mech-cad-design init --workspace W --actor A --organization O --design-group G
mech-cad-design status --workspace W
mech-cad-design migrate --workspace W [--dry-run]

mech-cad-design design start --workspace W --design-id ID --title T \
  --requirements-json '{"capacity": 4}' --proposal P --approve "yes"
mech-cad-design design list --workspace W
mech-cad-design design open --workspace W --design-id ID
mech-cad-design design status --workspace W --design-id ID
mech-cad-design design mistakes --workspace W --design-id ID

mech-cad-design gear size --power-kw 7.5 --pinion-rpm 1450 --gear-rpm 480 \
  --pinion-material 20MnCr5_carburised_G2 --gear-material 20MnCr5_carburised_G2 \
  --duty moderate --life-hours 20000 --safety-factor 1.5 [--out sizing.json]

mech-cad-design screening record --workspace W --design-id ID \
  --screening-file screening.json \
  --query-json '{"material_class": "linear_elastic"}'
mech-cad-design screening status --workspace W --design-id ID

mech-cad-design family start --workspace W --onboarding-id OB \
  --family-id F --family-name N [--alias A]
mech-cad-design family analyze --workspace W --onboarding-id OB --analysis-file P
mech-cad-design family review --workspace W --onboarding-id OB --decision "approved"
mech-cad-design family publish --workspace W --onboarding-id OB
mech-cad-design family status --workspace W --onboarding-id OB

mech-cad-design knowledge bootstrap --workspace W
mech-cad-design knowledge import-postgres --workspace W --source-env E
mech-cad-design standard-parts providers [--category C]

design mistakes reports which mandatory validation checks this design failed and later corrected, and which defects are still outstanding.

design start is idempotent: repeating it with the same design intent resumes the existing job instead of creating a second one. design open reports where an existing job stands and which step comes next. Starting a design needs a configured FreeCADCmd; listing, opening, and reading design jobs do not.

Select knowledge administration only when needed:

MECH_DESIGN_MCP_TOOL_PROFILE=knowledge-admin mech-cad-design-mcp

The current acceptance target is official FreeCAD 1.1.3. Configure the exact FreeCADCmd path and SHA-256 in the workspace or environment. Durable knowledge works out of the box on the local SQLite store; set MECH_DESIGN_DATABASE_URL to use PostgreSQL instead, which has no pgvector requirement. Install the neo4j extra (python -m pip install '.[neo4j]') only when the optional relationship projection is wanted.

Project-owned Agent Skills

• mech-cad-design

• freecad-standard-parts

• freecad-model-validation

Operating boundaries

• Generated models, reports, screenshots, databases, credentials, and customer-specific evidence stay outside the public repository.

• Local MCP and database services remain bound to loopback interfaces.

• A passed validation report proves only the checks that ran against one exact model revision. It is not FEA, manufacturing release, safety certification, or legal standards certification.

• A surrogate screening estimate is a learned prediction with declared coverage, not an analysis. It is not FEA either, it proves nothing about the model, and no screening result can complete, block, or confirm a design.

• Final engineering responsibility remains with the user or an authorized engineer.

Documentation

• Architecture and trust boundaries

• Surrogate screening

• FreeCAD GUI MCP integration

• Engineer learning playbook

• Database deployment

• Windows release acceptance

• Security policy

• Changelog

License

Project source is released under Apache-2.0. External dependencies, integrations, and assets retain their own licenses; see Third-Party Notices.