[Feature Request] First-class Windows CAD Computer Use for SOLIDWORKS and AutoCAD

What variant of Codex are you using?

Codex Desktop App on Windows (Computer Use, Skills, and MCP integrations)

What feature would you like to see?

Summary

Please add first-class Windows CAD support to Codex, starting with SOLIDWORKS and AutoCAD. The goal is a safe, CAD-aware agent interface that combines visual Computer Use with structured integrations such as MCP and COM where available.

Current limitation

I tested GPT-5.4 through GPT-5.6 with open-source Skills, MCP servers, automation scripts, and custom tooling.

Simple sketches, planar drawings, and basic extrusions can work, but reliability drops sharply when a real engineering workflow requires:

• multi-step feature-tree operations;
• dimensions and sketch constraints;
• assemblies and parametric relationships;
• stable selection of faces, edges, planes, and references;
• recovery after a failed or unexpected UI action;
• validation of the resulting geometry and native CAD file.

The current gap is not only model reasoning. Codex needs a reliable CAD execution and verification layer.

Requested capabilities

1. A safe Windows Computer Use bridge for CAD applications.
2. Structured SOLIDWORKS and AutoCAD integrations using supported APIs, COM, or MCP where possible, with UI automation as a fallback.
3. Awareness of the active document, selected entities, units, sketches, constraints, features, configurations, and assemblies.
4. Checkpoints, pause, replay, rollback, and resumable multi-step modeling workflows.
5. STEP, STL, DWG, and native-file export with geometry and file validation.
6. Clear progress and failure reports showing what succeeded, what failed, and why.
7. Local execution and explicit permission boundaries for proprietary engineering files.
8. A public CAD Computer Use benchmark so users can measure task completion and reliability.

Suggested acceptance criteria for an initial release

A useful first milestone could demonstrate that Codex can:

• open an existing CAD project and identify its current modeling state;
• create or modify a constrained sketch;
• create and edit parametric features without corrupting the feature tree;
• recover from one failed selection or command;
• save the native project and export a neutral format;
• verify that the output opens successfully and satisfies defined dimensional or geometric checks.

This request does not assume fully autonomous engineering or bypassing proprietary CAD licenses. The priority is a controlled, observable, and verifiable workflow for authorized local software and files.

Additional information

Relevant public dataset

The Markov CAD 1000 Hours dataset may provide a useful foundation for research, training, or evaluation:

https://huggingface.co/datasets/markov-ai/cad-1000-hours

Its dataset card reports:

• 1,021.64 hours of recorded CAD work;
• 597 workflows across 10 CAD, BIM, structural-analysis, and visualization applications;
• 8,973 artifact files and approximately 256.60 GiB;
• AutoCAD, SOLIDWORKS, CATIA, Siemens NX, Revit, SketchUp, and related applications;
• synchronized screen recordings, mouse and keyboard events, task descriptions, narration, rubrics, source files, and completed deliverables.

The dataset should be evaluated under its applicable license and usage terms. I am proposing it as a potential benchmark or research input, not as a requirement to redistribute or bundle it.

Why this matters

A reliable CAD workflow would extend Codex from software development into practical engineering design assistance for mechanical design, industrial automation, education, and maker workflows. Existing community Skills and MCP experiments show demand, but an official execution, safety, and validation layer is needed for dependable use.

No repository file, test, or entry point is named. Start by mapping the existing Windows Computer Use, Skills, and MCP integration boundaries; done would require an agreed initial milestone with acceptance tests for CAD state inspection, recovery, native-file saving, export, and geometry validation.