Zkitszo - The Real
October 21, 2026

Codex in the Driver's Seat of Fusion 360

A MO-RA 360 radiator is about the size of a car radiator. More often than not it ends up on the floor under a desk, where it needs feet. Watercool's own POM feet go for €25.90 at Caseking... for a pair of plastic blocks. Anyone with a 3D printer looks at that price and reaches for CAD. Anyone who has never opened CAD now has another option- let an AI agent drive Autodesk Fusion for them.

The agent can sketch, extrude, fillet and export by itself. Reading minds is not on its tool list though.. every "no, not like that" burns more tokens.. some of them a lot.

Getting the Agent Into Fusion-

The bridge is MCP, the Model Context Protocol- an open standard that lets an AI app call tools inside other programs (the MCP site pictures it as a USB-C port for AI apps). For Fusion the piece to grab is fusion360-mcp-server from Faust Machines. It's MIT licensed and still marked beta. It comes in two halves. An add-in runs inside Fusion and listens on localhost:9876. A small Python server sits between that socket and the agent, talking MCP over stdio- plain standard in/ standard out, the way any terminal program talks.

Ninety-three tools ship with it- sketches, extrudes, fillets, patterns, CAM, STL export, even a render_view that hands the agent a picture of the viewport so it can look at its own work. It's tested with Claude Code and lists Codex, Cursor and OpenCode as working clients.

How a prompt turns into geometry Planning chat any chatbot research, screenshots, writes the prompt Codex agent GPT-6 Luna gets finished prompts only MCP server fusion360-mcp-server launched by the agent via uvx Fusion360MCP add-in runs inside Fusion listens on localhost:9876 prompt stdio TCP Project folder AGENTS.md · HANDOVER.md TODO.md · REFERENCE.md read first, every new chat Fusion 360 sketches, features timeline (review here) units in centimetres

Fusion's API works in centimetres, so a 25 mm hole is 2.5 to the tools. One operation per tool call- batching crashes the add-in. The socket has no password at all, so it stays on localhost or a trusted home network. One tool, execute_code, runs any Python inside Fusion- keep that one on "ask before running".

Setting It Up

A German guide on Druckwerkstatt 3D walks the long install for Claude Desktop, virtual environment and all. The short route uses uvx, which fetches the server from PyPI and runs it- the repo is only needed for the add-in. Install uv first. macOS/ Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows PowerShell:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Grab the repo, step into it, then copy the add-in into Fusion's add-ins folder:

git clone https://github.com/faust-machines/fusion360-mcp-server.git
cd fusion360-mcp-server

macOS:

cp -r addon ~/Library/Application\ Support/Autodesk/Autodesk\ Fusion\ 360/API/AddIns/Fusion360MCP

Windows PowerShell:

Copy-Item -Recurse addon "$env:APPDATA\Autodesk\Autodesk Fusion 360\API\AddIns\Fusion360MCP"

In Fusion it's Shift+S → Add-Ins → Fusion360MCP → Run. The TEXT COMMANDS window should then report [MCP] Server listening on localhost:9876. Some Mac installs use a folder called Autodesk Fusion with no "360"- if the add-in is missing, copy it there too and quit Fusion properly (Cmd+Q).

Then Codex gets told about the server:

codex mcp add fusion360 -- uvx fusion360-mcp-server --mode socket

or the same entry by hand in ~/.codex/config.toml:

[mcp_servers.fusion360]
command = "uvx"
args = ["fusion360-mcp-server", "--mode", "socket"]

Claude Code takes the same line with claude in place of codex. Before any design work, ask the agent to call ping- a {"ok": true, "status": "pong"} back means both halves are talking.

Fusion's free personal-use licence covers home projects: non-commercial, individuals generating under $1,000 USD a year, fewer import/ export formats. Autodesk's page doesn't say either way whether add-ins are restricted on it.

Two Chats, One Agent

Every message to the agent is metered, every correction too. Talking a design through with it- "a bit wider.. no, the other wing.. not that wide"- is the most expensive way to use it. So the work splits across two windows. One ordinary chat does the thinking: research, dimensions, drawings, screenshots of the model so far, all the back and forth. Any chatbot will do, a free one included. Only a finished prompt goes to the agent, pasted in whole.

OpenAI's API price for GPT-6 Luna is $0.10 per million input tokens and $0.50 out. GPT-6 Astra, the flagship, is $10 and $50... a hundred times as much. Codex's Free plan lists Luna in the desktop app, still rolling out. A radiator foot doesn't need the flagship. More on stretching an agent's allowance in Usage Limit Reached? Not So Fast..

Luna, Astra and the Two Dials

Luna isn't Astra on a budget, though it gets pitched that way. OpenAI's own models page puts Astra at the top for complex reasoning and coding, GPT-6.1 Sol just under it, and Luna at the bottom as "our most efficient model for focused, high-volume tasks".

Codex also has two dials that get mixed up. The speed mode- Standard, Fast, Ultrafast- runs the same model quicker for more usage (Ultrafast burns included usage at eight times the Standard rate). Reasoning effort, from none up to max, does change the answer: OpenAI's reasoning guide says lower effort trades depth for speed and fewer tokens. "Turning it up only makes it faster" is true of the speed dial, wrong about the effort dial. Start Luna on low for a simple part. When the geometry gets fiddly, nudge the effort up before reaching for a bigger model.

The Project Folder

A long chat fills the context window- everything the model reads before it replies: our messages, its tool results, the files it opened. Once it's full, Codex compacts it, swapping the older turns for a summary. Whatever the summary dropped is gone, the "don't touch the mounting holes" from an hour ago included- from there the agent fills the gaps with guesses of its own.

Writing the memory down in files, one folder per project, keeps those rules out of the summary's reach:

  • AGENTS.md- the standing rules. Short answers, no saving unless asked, never change the part that bolts to the radiator, where the reference files live. Codex reads it before doing any work; /init writes a starter one.
  • HANDOVER.md- where things stand: which version is current, what got decided, what failed and why.
  • TODO.md- what's still left to do on the part.
  • REFERENCE.md- drawings and datasheets found so far, with the numbers pulled out. The hole spacing may take a second file or a pair of calipers.

When Codex reports that the context was compacted, it's time to hand over: the planning chat writes it, it goes into HANDOVER.md, a fresh chat starts with /new pointed at the folder. Pasting the old chat back in just refills the same window. A prompt that matters lives in a file- the same idea runs through Sloppy Jalopy- When a Trusted Prompt Goes Sideways.

Where It Goes Sideways..

Ask an agent to make a foot "more 3D-print friendly, lighter, still stable" and the assumptions roll in. Chat memory drags PETG over from some older project nobody mentioned this time. A prompt says start from V3 when V4 was saved minutes ago- nothing happens, tokens gone. A lightweighting sketch fails with "no closed profiles" because the agent cut that geometry itself two steps earlier. Name the material (PLA, PETG, whatever is on the spool), name the version, name what must not move.

Asked for a "honeycomb", an agent can paste a block of hexagon cut-outs into the part instead of reshaping it around a lattice. Undo, try again, same hexagons, round and round and round... Describing the structure works better than naming a pattern: two side ribs, two diagonals, a six millimetre outer edge, rounded corners carried all the way through. Then make the agent propose that geometry in words and wait for a yes- a plan is cheap to correct, fifteen minutes of Fusion surgery ending at two holes is not.

Answering a bad result with "I don't like it" gives the agent nothing concrete. It invents a new direction, then another one. Either say exactly what changes or roll back. Fusion's timeline, the strip along the bottom of the window, shows every step the agent took in order- the place to learn how it got built, the place to drag back to the last version that worked.

A prompt shaped like this, written in the planning chat, leaves the agent very little to guess:

Work in the folder water-cooling. Read AGENTS.md, HANDOVER.md and REFERENCE.md first.
Start from version V4 (already saved). Do not change the mounting area or the hole pattern.
Goal: lighter outer wings for FDM printing in PLA - less filament, same stiffness.
Before touching the model, describe the geometry you plan (ribs, widths, fillets) and wait for my OK.
One change per step. When done, save as V5 and update HANDOVER.md.

For all that, a free MO-RA3 360 foot already sits on Printables- standard M3 computer screws, PLA, 0.25 mm layers, four perimeters, 25% cubic infill, no supports, about 82 grams and two hours on the printer (CC BY-NC, so personal use). A whole night of prompting can end with printing that file anyway! The agent earns its keep on the step after that: taking a part that already exists and asking for the version with less plastic, a taller stance, a slot for a cable.

Same Trick, Other Programs

The MCP server has to match the program it drives. Blender has MCP for Blender, the project the Fusion server's socket bridge was modelled on. FreeCAD has freecad-mcp- same two halves, MIT licensed, no CAD licence strings at all. Good jobs for any of them: a TPU phone case from the phone's published dimensions, bins and hooks for a pegboard, a circuit board housing, a dust shoe for a CNC spindle.

Has anyone gotten an agent to design a part that printed right the first time? What did the prompt look like.. or did it end at the download button too? Leave a comment with the prompt- would love to see 'em.

Glossary

  • MCP (Model Context Protocol) — an open standard that lets AI apps call tools in other programs.
  • Add-in — a plug-in that runs inside Fusion; here, the half of the bridge that executes commands.
  • stdio — standard input/ output; how the agent and the MCP server pass messages, like a terminal program.
  • Socket / TCP port 9876 — the local network connection between the MCP server and the Fusion add-in.
  • uv / uvx — a Python tool manager; uvx downloads and runs a Python tool in one step.
  • PyPI — the public index of Python packages.
  • Codex — OpenAI's coding agent, available as a desktop app, a command-line tool and an IDE extension.
  • Token — a chunk of text, roughly part of a word; AI usage and prices are counted in tokens.
  • Context window — everything the model reads before it replies, up to a fixed limit.
  • Compaction — replacing older parts of a chat with a summary to free up the context window.
  • AGENTS.md — a Markdown file of standing instructions that coding agents read before starting.
  • Handover file — a written summary of where a project stands, used to start a fresh chat.
  • Reasoning effort — how much the model thinks before answering; higher means slower, more thorough.
  • Speed mode — Codex's Standard, Fast and Ultrafast settings; same model, faster output, more usage.
  • Timeline — Fusion's step-by-step history of every feature in a design, editable after the fact.
  • Closed profile — a fully enclosed sketch shape; Fusion needs one to extrude or cut.
  • Extrude / fillet — pulling a sketch into 3D / rounding an edge.
  • FDM — fused deposition modelling, the common filament 3D printing process.
  • PLA / PETG / TPU — common printing filaments: easy and stiff / tougher and more heat-tolerant / flexible.
  • POM — polyoxymethylene, also called acetal; a hard engineering plastic used for the official feet.
  • Infill / perimeters — the inner fill pattern of a print / the number of solid outer walls.
  • STL — the mesh file format slicers read.
  • CC BY-NC — a Creative Commons licence: credit the author, no commercial use.

Sources

Comments

Post a Comment

← Newer