Most AI demos end on a screen. This one ends with an object in your hand. Kiln now speaks the 2026-07-28 Model Context Protocol specification, and the part that matters to you is not the version number. It is that a design you asked for opens as a real 3D stage inside the conversation, that a slow job stops freezing the chat, and that nothing gets ordered without you actually saying yes.

The thing an AI assistant could never do: show you the object

Ask an assistant to design something and, until now, the best it could do was hand you a picture. A render is a flat opinion of a 3D thing. You cannot see whether the bottom sits flat, whether the wall behind the logo is thick enough, or whether the lid actually meets the box.

Now the object opens in the conversation. Drag to rotate it, scroll or pinch to zoom, tip it over and look underneath. It is the same 3D stage the Kiln web studio uses, running inline where you were already talking. Apps with interactive panels, Claude among them, show it right there. Every other client gets a link to the same stage, live for about 30 minutes.

That is MCP Apps, the interactive-interface extension to the protocol. Before it, a tool could return text, or an image, and that was the whole vocabulary. A 3D model is neither. Now a tool can hand back an interface, and 3D printing is a good argument for why that matters: the difference between reading that your part is 340mm tall and turning it over to see the overhang you are about to spend nine hours printing.

A print outlasts the conversation that started it

There is a scale problem buried in the phrase “long-running work.” For most agent tooling that means a minute or two of thinking, maybe a slow API. Here it means a machine extruding plastic in another room for the rest of the afternoon, and a job that ends in an object you have to walk over and pick up. No held-open connection was ever going to stretch across that, and pretending otherwise is how work gets lost at the very last step.

That is why the task handoff is the piece of the 2026-07-28 Model Context Protocol specification that actually matters for manufacturing. Kick the work off, take the ticket, and let the session end if it wants to. The job keeps running on Kiln’s servers and the finished result is waiting when you come back for it, instead of a chat that timed out somewhere around layer 400.

The same applies to the shorter geometry work that still outlives a comfortable request: carving a photo into a surface, repairing a broken mesh, generating an assembly booklet or a dimensioned drawing. In our own testing, a whole-object texture carve on a 340mm model handed back a ticket in under three seconds, finished in the background, and returned the completed part when asked for it. Nothing changes for clients on the older protocol; they get the same answer they always did, just synchronously.

Nothing gets ordered without you saying yes

Kiln can have a part professionally made and shipped to you, which means Kiln can spend your money. That deserves more care than a chat message.

Previously the assistant was asked to check with you before confirming a price or a shipping choice, and asking an AI to remember to ask is a promise, not a guarantee. Now the confirmation is part of the connection itself: the price comes back as a real question in your app, the shipping choice comes back as a real question, and nothing is confirmed until you answer. The protocol carries the request for consent, so it does not depend on the assistant's good behavior.

Why a 3D printing company cares this much about a protocol

Because the last mile of AI is not text. It is the moment something leaves the screen.

An assistant that writes you a poem is finished when the poem appears. An assistant that designs you a bracket is not finished until the bracket exists, fits, and holds. Everything between those two points is unglamorous work: is this printable, will it warp, does the hole come out the size the drawing says, which of my machines should run it, how do I fix it when it fails at layer 400. Kiln is that middle. The protocol is how your assistant reaches it.

So a richer protocol is not a footnote for us. An interface a tool can hand back is the difference between describing an object and showing it. A ticket for slow work is the difference between a chat that survives real geometry and one that times out. A confirmation built into the wire is the difference between hoping an assistant checks with you and knowing it must.

Trying it

The hosted connector needs no install. Point any MCP client at Kiln's endpoint with your key and the curated server-side surface is there: design, texture, engineering math, material and printability intelligence, cutting an oversize model into printable pieces, and ordering. Tools that drive a physical printer are deliberately not on that surface, because they need to run on the machine your printer is plugged into. For those, install Kiln locally.

Claude Code: claude plugin marketplace add codeofaxel/Kiln, then claude plugin install kiln@kiln.
Everything else: pip install kiln3d and follow the install guide.

Kiln is open source, and the protocol work is in the open with it. If you want the full list of what landed alongside this, the 1.3.0 release notes cover it.

Get Kiln →