How to Connect Gemini to AutoCAD (MCP, 2026)

Search for this and the top result is a connector directory with no traffic to its name, followed by four videos and somebody's LinkedIn post. Nobody has written down how it actually works, which is a shame, because it does work and it is the cheapest version of this in the whole family. The Gemini app on the web cannot touch your drawing and never will. Gemini CLI, which is a different product with the same name on it, can. Here is the wiring, what it costs, the one setting that will quietly hand a language model the keys to your machine, and the jobs it still cannot do.

FreeUpload a photo · freeGet the drawing a chatbot cannot give youPhotograph the part and get real geometry with real dimensions, not a picture of a drawing. Your first drawing is free, no card needed.Drop a photo here, or click to choose. JPG, PNG or HEIC.

The short answer

Install Gemini CLI, install an AutoCAD MCP server, tell the CLI where the server is, and Gemini can draw in your open AutoCAD session. The whole trick is that Gemini CLIis a terminal application, and terminal applications can start other processes on your machine. The Model Context Protocol is the standard that lets a model call tools those processes expose. The AutoCAD server exposes eight of them, covering drawings, entities, layers, blocks, annotation, P&ID symbols, views and system operations, plus a tool that runs arbitrary AutoLISP. Gemini picks the tool, the server does the drawing.

Two different things get called this

Before any of the setup, the disambiguation that will save you an hour. There are two Google products that connect Gemini and Autodesk, they do completely unrelated jobs, and search results mix them freely.

If what you want is Gemini answering “how do I set up a named page setup” out of Autodesk's own manuals, the Autodesk Product Help connector is the right answer and the rest of this article is not. Everything below is about the second column.

What this actually costs

This is the part that makes the Gemini route interesting rather than just another option, so it is worth doing the arithmetic properly.

• Gemini CLI: free. Apache 2.0, and signing in with a personal Google account gives you 60 requests per minute and 1,000 per day against Gemini 3 models with a one million token context window. A drafting session is nowhere near a thousand requests.
• The AutoCAD MCP server: free. MIT licence, so you can read it, change it and ship things built on it.
• AutoCAD: not free, and this is the only real cost. But LT is the cheap tier and LT works, which matters more than it sounds.
• Or skip AutoCAD entirely. The server has a headless backend that writes DXF with no AutoCAD installed. You lose the live application, you keep the file.

Compare that with the other two routes we have written up. Connecting Claude to AutoCAD and connecting GPT-6 Astra to AutoCAD both assume a paid subscription on the model side. Gemini does not. If you are trying this out to see whether it is useful at all, that is the difference between an experiment and a purchase.

What you actually need

The requirements are stricter than the videos suggest, and they split depending on which backend you run.

Two things worth saying plainly because they trip people up. LT 2024 really is enough, since AutoLISP only arrived in LT with that release and this server drives LISP rather than loading a .NET plugin. And Mac users are not completely stuck: AutoCAD LT for Mac has no AutoLISP so the live path is closed, but the headless backend will still write you a DXF. The wider survey of what is out there, including the FreeCAD servers, is in CAD MCP servers.

Wiring it up

Four steps, and the third is the only one with a trap in it.

1. Install Gemini CLI. npm install -g @google/gemini-cli, or brew install gemini-cli, or run it once with npx @google/gemini-cli. Sign in with your Google account on first launch.
2. Install the AutoCAD MCP server. Clone it, create its virtual environment, and note the full path to the Python executable inside that environment. You will need the path, not the command name.
3. Register the server. Either run gemini mcp add or write the config by hand. The command form is gemini mcp add [options] <name> <commandOrUrl>, and the option that matters is -s, which chooses the scope. It defaults to project, meaning the server is registered only for the folder you are standing in. Pass -s user if you want it available everywhere, which for a CAD server you almost certainly do.
4. Start AutoCAD, then start Gemini. The live backend needs the application already open.

Written by hand, the config lives at ~/.gemini/settings.json for the user scope or .gemini/settings.json inside a project, and it goes in an mcpServers object. Each entry needs one of command, url or httpUrl, and then takes optional args, env, cwd, timeout, trust, includeTools and excludeTools. A local server started as a process uses command with args of ["-m", "autocad_mcp"], and the backend is selected with an AUTOCAD_MCP_BACKEND environment variable set to auto, file_ipc or ezdxf.

The setting that removes every safety prompt

There is one option in that list that deserves its own section. Setting trust to true on a server, or passing --trust when you add it, bypasses every tool call confirmation for that server. Gemini stops asking. It just does it.

For a read-only server that fetches documentation, fine. For this one, stop and think. Among its eight tools is one that executes arbitrary AutoLISP, and AutoLISP is a real programming language with access to your file system. Trusting this server means a language model can run code it wrote itself, on your machine, over your drawings, with nothing in between. The failure mode is not malice, it is a confident model misreading an instruction and running something destructive on the file you had open.

Leave trust off. The confirmations are mildly annoying for about a day and then you stop noticing them, and the first time one catches something you will be glad. If you do turn it on, work on a copy of the drawing and keep it out of anything shared.

What it is genuinely good at

Having run the other two models at this, the pattern is the same and it is narrower than the demos imply. This is good at geometry you can describe precisely and would find tedious to place by hand.

• Patterned and parametric work. A bolt circle, a grid of holes, a run of identical brackets at a spacing. You say the rule, it places the entities.
• Layer and block housekeeping. Renaming, moving entities between layers, freezing, inserting blocks consistently. Dull, rule-shaped, easy to verify.
• Writing the LISP you would have looked up. Gemini is strong at code, and AutoLISP has decades of public examples. Ask for the routine, read it, then run it.
• Batch edits with a clear rule. Anything you could describe to a junior drafter in one sentence.

If prompting is what you are really after rather than wiring, Gemini prompts for technical drawing is the prompt-side version of this, and can Gemini make technical drawings is the honest capability test with the output shown.

What it cannot do

The limits are worth knowing before you build a workflow on this, and most of them are not really about Gemini.

It cannot turn an image or a PDF into a DWG.This is the first thing people try and the first thing that fails. There is a thread on Google's own support forum from someone trying exactly this on Gemini 1.5 Pro and finding the accuracy insufficient. That is an older model now, and it does not matter, because the reason is not one a newer version fixes. Reading geometry out of a picture is a measurement problem. A photograph stores no scale, so nothing in the chain knows whether the part is 50 mm or 500 mm across. That is a different tool: the route from a picture to a real file is in image to DWG, and the version of it that goes through a chatbot first is converting an image to CAD with ChatGPT.

It cannot invent a dimension. Same reason. Give it one real measurement and everything else follows; give it none and every number it writes is a plausible guess. That idea in full is how to get dimensions from a photo.

It cannot do 3D in LT, because LT cannot. Nor ActiveX or COM automation, nor the Visual LISP IDE, which quietly rules out a number of tutorials still circulating. And the P&ID tools need a specific third-party symbol library installed at a specific path, so treat that part as optional rather than included.

Before you trust what it drew

One habit worth building. Whatever the model reports back, check the drawing rather than the transcript. Models describe what they intended to do, and the description is usually right even when the geometry is not. Measure one feature you specified, confirm the entity landed on the layer you asked for, and confirm nothing else moved while it worked.

On a fabricated part a wrong number is expensive in a way an incorrect sentence never is, which is the whole argument for looking. If you are handing the result to a shop, our DXF checker will tell you whether the file is actually cuttable before anyone quotes it, and why your DXF will not cut covers the failures that look fine on screen.