The most asked question from the first Fusion video was "will this work with my account?" The answer is in one place. In Fusion 360, click your initials in the top right corner, open Preferences, then click API. Look for "Fusion MCP server" with the line "runs locally on this device." If you can check that box, it will work, whatever account type you are on. If it is grayed out or missing, your account does not have it yet.
Two requirements beyond the checkbox: you need the Claude desktop app, and Fusion 360 has to be open while Claude works. Fusion itself is a subscription product; the MCP connection does not cost extra on top of it.
The test was simple to describe and honest by design. Take a screenshot of an existing sketch with its dimensions, hand it to Claude with "create this in Fusion 360," and add no other help. No file, no measurements typed in, no coaching.
About five minutes later there was a finished 3D part sitting in Fusion. It looked right, but looking right is not the test. Pulling the measurements is: the section that should be 110 millimeters measured 110 millimeters, and the 180 millimeter vertical came back 180. It even placed the text from the drawing. From a screenshot.
The honest caveat: a CAD engineer builds this part in a few minutes without help, and that is fine. This is not a tool for the diehard engineer. It is a starting line for someone who has an idea, does not yet know a sketch from an extrude, and wants a base to learn from instead of a blank screen.
The second design went in the same way, but its screenshot carried much less to go on. The result was a partial build, a few cylinders roughly where the part should be, and a follow-up "complete everything" did not rescue it. Two attempts, both short of the target.
That failure is the most useful data point in the whole test. The model did not get worse between part one and part two. The input did. When the screenshot carried full dimensions, the build was spot on; when it did not, the build stalled. If you try this, the quality of what you feed it is the dial that matters.
| Task | Result | Usage burned (5x plan, 5-hour window) |
|---|---|---|
| Part from a detailed screenshot | Success, dimensions verified | 13% |
| Part from a thin screenshot | Failed, partial geometry | 7% |
| "Complete everything" retry | Still incomplete | 7% |
So yes, it is token heavy. One good part for roughly an eighth of a usage window is workable for a hobbyist or a prototype; it is not a production pipeline. Knowing the number before you start is the point of publishing it.
Strip away the CAD and this is a repeatable lesson about every AI tool you will ever evaluate: test it on your real work, verify the output against something measurable, note what it consumed, and let it fail in front of you before you rely on it. That is the same method behind everything KiS does, including connecting AI to your own business records: the tool is only as good as the information you hand it.
The free quiz takes 2 minutes and gives you a starting list for your business. No call, no card.
Take the free 2-minute quiz