CAD-MCP is an MCP server that lets Claude or another MCP-compatible client control CAD software installed on Windows through the COM interface. The daobataotie/CAD-MCP repository states support for AutoCAD, GstarCAD, and ZWCAD, with basic drawing, text, hatch, dimension, layer, and DWG save functions. Unlike AutoCAD MCP, it does not offer a headless DXF backend: it assumes CAD software is installed and reachable through Windows.
For a firm, CAD-MCP reads as a simple connector for automating repetitive 2D actions in a DWG-compatible environment. It is less ambitious than Revit MCP, less feature-rich than AutoCAD MCP, but more direct: Python, pywin32, a config.json file, then MCP tool calls that create lines, circles, arcs, polylines, rectangles, text, hatches, and dimensions. That simplicity is useful, as long as it is not confused with a complete CAD platform.
Contents
- What CAD-MCP does
- Windows COM architecture
- Useful capabilities for a firm
- Comparison with AutoCAD MCP and Revit MCP
- Security and limits
- Educasium's position
- Sources and next reading
What CAD-MCP does
Key point: CAD-MCP turns structured or natural-language commands into simple 2D CAD operations. The README presents it as a CAD control service driven by natural-language instructions, with support for AutoCAD, GstarCAD, and ZWCAD.
Exposed tools
The README lists the main functions as drawing a line, circle, arc, polyline, rectangle, text, hatch, dimension annotation, saving the drawing, and processing a natural-language command. The verified server.py file also exposes draw_ellipse, which slightly expands the real surface beyond the short README list. The verified set should therefore be treated as a small drawing toolkit, not an exhaustive CAD API.
Operations create entities in the active document's ModelSpace and can receive parameters such as layer, color, or lineweight depending on the tool. The controller creates the layer if it does not exist, then assigns the entity to it. That is practical for simple drawing generation, but it assumes the firm defines its layer conventions beforehand.
Natural-language processing
The project includes an nlp_processor intended to convert text instructions into CAD parameters. The README mentions color recognition, shape keyword mapping, and action keyword mapping. This layer can make the demonstration more accessible, but it does not replace professional validation.
In a firm, good usage means keeping instructions tightly scoped: coordinates, entity type, layer, color, text height, and save path. The vaguer the instruction, the higher the risk of producing a syntactically valid but graphically wrong drawing. Natural language helps the interaction; it does not remove the need for specification.
Windows COM architecture
Key point: CAD-MCP depends on Windows, pywin32, and installed CAD software. It is not a universal web service drawing in the cloud.
CAD software connection
The verified cad_controller.py imports win32com.client and pythoncom, then selects an application identifier according to the configured CAD type: AutoCAD.Application for AutoCAD, GCAD.Application for GCAD or GstarCAD, and ZWCAD.Application for ZWCAD. The controller first tries to connect to an existing CAD instance, then tries to launch a new instance if none is found.
This approach matches Windows desktop automation: the server controls a local, visible application installed on the workstation. It is not suited to a Linux server, CI environment, or machine without a CAD license. The verified requirements.txt dependencies are pywin32, mcp, pydantic, and typing.
Configuration and startup
The README says configuration lives in src/config.json, including CAD software type, startup wait time, and command delay. Startup uses Python on src/server.py, then the MCP client configuration points to that script. This is more manual than a ready-to-run npm package.
That simplicity may be enough for a test workstation. However, a firm that wants regular usage needs to stabilize the Python path, CAD type, output folder, and installed versions. Without that, two workstations may run the same instruction with different behavior.
Useful capabilities for a firm
Key point: CAD-MCP is relevant for bounded 2D drawings, simple templates, and automation demonstrations. It should not be presented as a BIM assistant or a replacement for a CAD technician.
Drawing and annotation
Drawing functions cover classic primitives: lines, circles, arcs, ellipses, polylines, rectangles, text, hatches, and linear dimensions. This is enough to build a sample plan, simple grid, principle diagram, or repetitive geometry. Coordinate parameters remain essential, though: the assistant does not automatically know the full drawing context.
Available dimensioning is linear or aligned according to the controller. For a professional deliverable, this remains insufficient without graphic verification: dimension styles, scales, units, text, orientation, and title block conventions are not treated as a full system in the README. Automation can place elements, not guarantee a ready-to-deliver drawing.
Layers and DWG saving
Layer support is more interesting for firm usage. The controller can create a layer, make it active, and assign entities to it. For generating repetitive elements within a DWG convention, that is a useful base, provided the convention is explicitly supplied.
Saving the drawing goes through SaveAs to a specified path. This capability needs governance: on a production workstation, a poorly targeted save command can overwrite a file or create a version in the wrong place. As with any automated CAD workflow, start in an isolated test folder.
Comparison with AutoCAD MCP and Revit MCP
Key point: CAD-MCP is the most general CAD connector, but not the most complete. Its strength is multi-CAD support through COM; its limit is functional depth.
| Option | Natural use | Main strength | Limit to watch |
|---|---|---|---|
| CAD-MCP | Simple 2D drawing in AutoCAD, GstarCAD, or ZWCAD | Same MCP logic for several Windows CAD tools | No headless backend and limited surface |
| AutoCAD MCP | Automating AutoCAD LT and producing DXF | Two backends, 8 consolidated tools, ezdxf | More specific to AutoCAD LT and DXF |
| Revit MCP | Acting on a BIM model | BIM data, levels, rooms, families | Does not cover 2D DWG flows |
CAD-MCP versus AutoCAD MCP
The AutoCAD MCP is richer for AutoCAD LT: it distinguishes File IPC and ezdxf, documents operations by backend, and covers functions such as blocks, P&ID, screenshots, and execute_lisp. CAD-MCP is simpler and more general across compatible software, but it does not provide the same level of detail about per-operation limits.
The choice therefore depends on the software fleet. If the firm works strictly in AutoCAD LT 2024 or newer and also wants to generate DXF without AutoCAD, AutoCAD MCP looks more suitable. If the team uses GstarCAD or ZWCAD and wants to test basic natural control, CAD-MCP is worth studying.
CAD-MCP versus Revit MCP
The Revit MCP acts on a BIM model, with element, level, room, and model-data logic. CAD-MCP acts on 2D drawing primitives in a CAD model space. The risks and gains are not the same.
For an architecture firm, CAD-MCP makes sense in detail DWG, revision, or diagram workflows. It is much less relevant for structuring a model, extracting BIM quantities, or checking professional parameters. The best use respects the real place of DWG in the firm.
Security and limits
Key point: CAD-MCP can modify the active CAD document; it should therefore be tested on copies. Even a simple command can produce a wrong drawing if coordinates, layers, or units are poorly defined.
Windows workstation dependency
The dependency on pywin32 and COM makes the project highly local. It needs a Windows workstation, the right CAD software, a correct src/config.json configuration, and a session able to launch or control the application. This is not a connector to deploy indiscriminately on a server.
The controller tries to connect to an existing instance, then may launch a new one. That behavior can surprise users who think they are working in a specific drawing. Before every test, verify the active document and save path.
Functional limits
The README talks about basic drawing, layers, and saving. Code comments indicate that functions such as entity deletion, move, rotate, and scale were planned for later. Even though the controller already contains useful primitives, a complete CAD operation surface should not be claimed.
The NLP layer is also a limit. A natural sentence can be misinterpreted, especially when it contains implicit measurements, graphic references, or internal conventions. For professional usage, favor structured prompts and review the result in the CAD software.
Educasium's position
Key point: CAD-MCP is a good teaching support for explaining CAD automation, not a production reference already validated by Educasium.
Available experience
Educasium has not yet deployed CAD-MCP in a client firm or measured verified time savings on DWG drawings or GstarCAD/ZWCAD. This analysis relies on the README, requirements.txt, server.py, and cad_controller.py verified through GitHub. That precision prevents a technical review from turning into a commercial promise.
Its teaching value is real. CAD-MCP shows how an assistant can call structured CAD tools, how a Windows application can be controlled through COM, and why the quality of a technical instruction matters as much as the AI model used. For training, it is simpler to explain than a full BIM connector.
Hypothetical implementation
Hypothetical implementation. For a pilot, Educasium would start on an isolated Windows workstation with a test DWG file, explicit CAD configuration, and a separate output folder. First tests would cover draw_line, draw_rectangle, draw_text, add_dimension, and save_drawing, with a minimal layer convention supplied in the prompt.
The second step would test process_command, but only with simple, verifiable instructions. If the team uses GstarCAD or ZWCAD, the COM identifier should be checked separately on the installed version. The expected result would not be a final drawing, but a list of repetitive CAD tasks the assistant can execute without degrading graphic quality.
Sources and next reading
Key point: the information on this page comes from the GitHub repository and source files verified on August 26, 2026. Multi-CAD claims come from the README and the application identifiers selected in cad_controller.py.
What to read next
If your firm mostly works in AutoCAD LT or DXF, also read the AutoCAD MCP. If your workflow is BIM, compare with the Revit MCP and Archicad MCP.
To place these tools in a broader strategy, read the AI software comparison for architects and the Mastering AI in Architecture course. The topic is not automation for its own sake, but choosing the tasks where AI control truly reduces friction.