Three ways to connect, compared
| In-process plugin (AutoGAD) | External automation (COM / Claude Code) | Chat with an export | |
|---|---|---|---|
| Where the model lives | inside AutoCAD, docked palette | outside: a Python script or an agent such as Claude Code | a browser tab |
| What it sees | live drawing database via tools | whatever your script reads through the ActiveX API | the DXF / text you paste |
| Edits | confirmed write tools, one Undo per call | anything the script does, no confirmation layer unless you write one | none, you retype |
| AutoCAD versions | 2025 / 2026 (.NET 8) | any version with COM, including old ones and via accoreconsole |
any |
| Setup | installer, paste key | Python + pywin32, your own prompt and tool loop |
none |
| Good for | daily engineering work on real drawings | batch jobs, one-off migrations, experiments | quick questions about a fragment |
1. A plugin inside AutoCAD
This is what I use every day and what AutoGAD is: a .NET 8 DLL that adds a ribbon tab and a chat palette. Every conversation starts with a compact snapshot of the open drawing (units, extents, layers, layouts, entity counts, block inventory, table text). When the model needs more it calls read tools over the live database — query_entities, get_block_inventory, get_layers, get_tables — and when it wants to change something it calls write tools that show you the exact change first: replace_text, set_text, create_text, set_layer_state, set_custom_property, run_command.
The key is stored DPAPI-encrypted, the provider can be Anthropic or any OpenAI-compatible endpoint, and each answer ends with its token cost. Installation is five steps and needs no admin rights; they are listed on the project page.
2. Driving AutoCAD from outside: Python, COM, Claude Code
AutoCAD has exposed a COM/ActiveX automation API for decades. From Python it is a few lines with pywin32: win32com.client.Dispatch("AutoCAD.Application"), then ActiveDocument.ModelSpace to iterate entities or SendCommand to feed the command line. An agent that can run code — Claude Code on your machine, or your own loop over the Anthropic API — can therefore control a running AutoCAD without any plugin.
What you get: it works on any AutoCAD version, it is easy to prototype, and it is the natural route for batch jobs over many files (with accoreconsole.exe for headless runs). What you give up: every call crosses a process boundary and is slow; there is no confirmation layer unless you write one; the agent sees what your script chooses to serialise, not the drawing; and there is no memory between sessions unless you build it. If you want Claude Code or Claude Desktop to talk to AutoCAD properly, the cleanest bridge is an MCP server that exposes the same read and write tools AutoGAD has — the tool list above is a workable spec.
3. Chat with a DXF
DWG is binary and a chat window cannot open it. DXF is text, so a small drawing or a fragment pasted into the chat is readable, and Claude reasons about it well: counts, inconsistencies, what a block contains. The limits are size (a real project drawing is far larger than a context window) and the fact that every change has to be applied by hand. Good for a question, wrong for a workflow. The details are in Does Claude work with AutoCAD?.
Can Claude draw in AutoCAD?
Two different questions hide here.
Can it create geometry? Yes, in a limited and useful way. Claude can write a DXF file as text, an AutoCAD script (.scr) that runs LINE, CIRCLE, PLINE with coordinates, or an ezdxf Python script — and AutoCAD opens the result. For a schematic sketch, a frame, a grid of fixtures with known spacing, this works. Inside AutoGAD the write tools are text-oriented; run_command can queue non-interactive command lines, but the plugin is built to read, check and edit real drawings, not to draft them.
Can it draw from a picture? Not a real drawing. Claude reads images well and can turn a sketch or a photo of a scheme into a rough DXF, but it estimates dimensions, has no notion of your layers, blocks and standards, and produces something a drafter then redraws. AutoGAD has no image input at all; its FreeCAD sibling FreeGAD does (take_screenshot), and even there the model uses the picture to check its own edits, not to trace.
What the model is genuinely good at is the other direction: read a finished drawing and audit it — count blocks against the schedule, find text on wrong layers, replace a project code everywhere, size a breaker from the loads it just read.
AutoCAD 2025 vs 2026: does it matter for AI plugins?
For plugins, no. AutoCAD 2025 is release R25.0 and AutoCAD 2026 is R25.1; both run plugins on .NET 8 with one binary-compatible managed API, so a single AutoGAD.dll loads in both. The installer registers the plugin under every R25.x profile of your Windows account. AutoCAD 2024 and earlier host plugins on .NET Framework and need a separate build, which AutoGAD does not ship. AutoCAD LT has no .NET API at all, so the plugin route is closed there; COM automation is the fallback.
One practical difference: after a per-machine MSI install, the Autodesk bundle loader only works when the system variable APPAUTOLOAD is not 0. The per-user installer's registry demand-load works regardless.
Keys, models, cost
The plugin is free (AGPL-3.0); you pay your model provider per token. For Claude, create a key in the Anthropic console. The default model is claude-opus-5; any current Claude model works, as does any OpenAI-compatible endpoint — OpenAI, OpenRouter model ids, or a local server if the drawing must not leave your network. A typical answer costs cents because the drawing snapshot is prompt-cached: 3 calls · in 1,234 (+18,000 cached) · out 800 · ≈ $0.05 this turn.
Which one should you pick?
- You design in AutoCAD daily and want the assistant next to the drawing: the plugin.
- You have a hundred DWGs to process once, or an old AutoCAD: COM automation from Python or Claude Code.
- You have one question about one fragment: paste the DXF into chat.
If your CAD is neither of these, or your company has its own drawing standards, the pattern transfers: snapshot, read tools, confirmed writes, memory. I build that live on a shared screen — see hire a CAD plugin developer.