Two developments a few weeks apart have done something the on-demand manufacturing industry has talked about for a decade and never shipped: they made the path from design intent to a physical part addressable by software, end to end, without a human in the middle of every step.
Autodesk released official Model Context Protocol support for Fusion in May, letting external AI agents take action inside a live CAD session. In July it followed with a cloud-hosted Fusion Data MCP covering project structure, folder management, and file permissions, and it has previewed a Revit server on the same architecture. Separately, SendCutSend closed a $110 million round on the strength of a business that automates quoting, nesting, and order routing so completely that orders reach the shop floor within seconds of checkout.
Neither announcement is dramatic on its own. Together they close a loop.
What Autodesk actually conceded
The text-to-CAD demos accompanying the Fusion MCP launch are the least important part of it. The significant move is that a tier-one CAD vendor published a vendor-blessed protocol allowing third-party AI to operate inside its application.
That is a public admission that the AI interface layer for engineering software will not be built in-house. Autodesk is betting it is better off being the kernel that executes than the assistant that reasons, and it is betting that customers will arrive with an agent they already chose.
The pressure now sits with Siemens, Dassault Systèmes, and PTC. Each faces the same fork: open the platform to external agents and accept a commoditised interface layer, or build a competing agent and try to keep customers inside it. The Revit preview suggests Autodesk considers the question settled across its portfolio rather than specific to Fusion.
Community developers got there first and still lead on coverage. Several independent MCP servers run as Fusion add-ins, exposing sketch creation, parameter manipulation, generic API calls, and in some implementations arbitrary Python execution inside the Fusion runtime. The official server is narrower and better supported.
The ordering side was already waiting
Fulfilment needed the least work to become agent-ready, because instant quoting engines have been production infrastructure for years. Upload a DXF or STEP file, pick material and thickness, add bending, tapping, hardware, deburring, or powder coat, and get price and lead time back in seconds.
SendCutSend built that as custom in-house software from launch, which is why it describes itself as a software company that happens to own lasers, and why the customer base runs from hobbyists to a large slice of the Fortune 500, private space firms, and defense contractors. The $110 million validates the model rather than the machinery.
Xometry, running the opposite bet as a broker across a network of vetted shops, has gone a step further on machine access. Its developer documentation now carries an llms.txt index written explicitly for AI agents, pointing at Markdown pages and OpenAPI endpoints. Companies do not publish agent indexes for their APIs unless they expect agent traffic to matter. Xometry grew revenue 36 percent year over year in the first quarter and has extended auto-quoting into injection molding across US and European markets.
Two models, and only one survives a commoditised interface
The distinction between these two companies is usually collapsed and should not be. Xometry is capital-light brokerage where the moat is the routing algorithm and the supplier graph, and where quality consistency across hundreds of shops is the standing problem. SendCutSend is capital-heavy vertical integration with a deliberately narrow process menu, mostly sheet and plate, no general CNC machining beyond 2D routing, where capability is capped but quality is controllable.
A unified agent interface is indifferent between them, and that indifference is the threat. If quoting becomes a commodity endpoint that agents poll across providers, the integrated fabricator’s software advantage stops being a moat and becomes something its customers now get elsewhere. What remains is price and turnaround, a considerably harder business than the one that raised nine figures.
The gate that has not moved
Design for manufacturability review is still human at both companies. Applications teams check part files after the order is placed and reach out when geometry will not manufacture. Autodesk’s own framing around the MCP launch makes the same point: manufacturability requires engineering rigor and domain expertise that lives in the kernel and on the shop floor, not in the model.
That gate exists because hardware inverts the economics that made agentic coding work. A coding agent iterates in milliseconds against free failures. A hardware agent iterates in days against failures that cost material, machine time, and freight. Brute force is not available.
The signal to watch is not another protocol announcement. It is whether manufacturability validation moves from post-order review to a pre-order API response. Until an agent can learn its part is unbuildable before money moves, every pipeline assembled from these pieces still has a person standing in the middle of it.
Leave a Reply