The short answer
If you upload a .dwg to a chat window, the model gets a binary blob it cannot parse. You can paste a DXF excerpt, a screenshot or a text export, and Claude will reason about it well — but it is reasoning about a copy, not your drawing. Every change still has to be typed back by hand.
The useful configuration is the reverse: the model lives inside AutoCAD and reads the live drawing database through the managed API. I built that plugin — AutoGAD — because I spend my days producing electrical and lighting working documentation in AutoCAD and wanted the assistant next to the drawing, not in a browser tab. This note is what it actually does, so you can decide whether the same pattern fits your work.
What the model sees
Every conversation starts with a compact snapshot of the active drawing: metadata, units, extents, layers, layouts, entity counts, the block inventory and the text of schedules. The snapshot is deliberately small and stable so that prompt caching works and repeated questions are cheap.
When it needs more, the model calls read tools over the live DWG database: query_entities, get_block_inventory, get_layers, get_tables and get_context. So "how many VPD projectors and SKR linear fixtures are installed?" is answered by counting real block references, not by guessing from a picture.
What the model can change
Write tools: replace_text (global find and replace across text, attributes, MLeaders and table cells), set_text, create_text, set_layer_state, set_custom_property and run_command, which hands an arbitrary command line to AutoCAD.
Every write opens a confirmation dialog with the exact change before anything happens. There is an Auto-approve toggle and an "Allow all this session" button for when you trust the session — off by default. A typical confirmed request: "Replace the project code К2.АХП-2.2025 with К2.АХП-3.2025 everywhere."
What it does not do
- It does not draft geometry for you from a description. It can run commands and edit what exists; a floor plan still comes from a drafter.
- It does not write into the DWG behind your back. Notes and chat history live in
%APPDATA%\AutoGAD, keyed by a fingerprint GUID of the drawing, so they survive renames and moves without touching the file. - It does not run on AutoCAD 2024 or older. Those versions host plugins on .NET Framework; AutoCAD 2025 moved to .NET 8, which is what AutoGAD targets.
Memory across sessions
remember stores per-user notes ("I always size cables for 380 V and cos φ 0.95") and per-drawing notes (project code, which layers and blocks mean what, decisions made). They are injected into later sessions, so the assistant stops asking the same questions. AUTOGADMEM shows and deletes what it remembers.
Models and cost
The default model is claude-opus-5 through the Anthropic API; any OpenAI-compatible endpoint works too — OpenAI, OpenRouter or a local server, which also answers the "can this stay offline?" question. The client is raw HTTP, no SDK, because a plugin hosted inside AutoCAD cannot carry one.
Each answer ends with a cost line, for example 3 calls · in 1,234 (+18,000 cached) · out 800 · ≈ $0.05 this turn, $0.21 this session. The cached figure is the drawing snapshot being reused between turns. The API key is stored with Windows DPAPI, so it never sits in a config file in plain text.
Installing it
A per-user Inno Setup installer puts the plugin under %APPDATA%\AutoGAD\bin and registers it for demand-loading in every AutoCAD 2025/2026 profile of your account — no admin rights. A per-machine WiX MSI with an Autodesk Autoloader bundle exists for GPO or SCCM deployment. Start AutoCAD, an AutoGAD ribbon tab appears, AUTOGAD opens the chat palette. The current version is 1.0.5.
The same thing for FreeCAD
FreeGAD is the sister project: the same agent as a FreeCAD 1.0/1.1 workbench in Python, with one extra tool — take_screenshot sends the rendered 3D view to the model so it can check its own edits visually. The design notes for both plugins are in a separate write-up.