NX CAD MCP: The Short Version
NX CAD MCP means giving an AI agent typed tools, over the Model Context Protocol, to read and build geometry in Siemens NX through NXOpen. inteliScience builds these servers for engineering teams: read tools first, every write gated by preview and confirmation. For a manufacturer in Asia we shipped a chatbot that inspects open parts and builds features on request.
| Chat assistant without MCP | NX CAD MCP assistant | |
|---|---|---|
| Sees the open part | no — you paste feature names by hand | yes — reads bodies, features, expressions, mass properties |
| Changes geometry | explains which command to click | creates sketches, extrudes, holes, patterns after confirmation |
| Trust model | none needed, nothing happens | preview, confirm, undo mark, audit log |
| First value | onboarding questions | repetitive parametric work, part inspection, variant builds |
The Problem: A Chatbot That Cannot See the Part
Every NX team that tried a general chatbot hit the same wall within a week. The assistant explains the Extrude dialog beautifully, but it has no idea that the part on your screen has 214 features, three suppressed bodies and an expression called p_flange_t that drives half the model. So the engineer describes the part in prose, the model answers in prose, and the engineer still does all the clicking.
That gap is not a language-model problem. It is a connectivity problem, and since Anthropic introduced MCP in November 2024 there is a standard way to close it. MCP has since moved to the Linux Foundation’s Agentic AI Foundation (December 2025), and its SDKs see close to half a billion downloads a month as of the July 2026 specification release. The protocol is no longer exotic; what is still rare is a CAD server behind it that an engineering department can trust.
We covered what a Siemens NX MCP server is at the architecture level on our service page, and the read-only, knowledge-grounding variant in our piece on NXOpen MCP. This article is about the geometry side: what an agent can actually read and build in NX today, how we structured the tools in a project we delivered this year, and where the open-source servers stand as of September 2026.
If your team works in Claude Code, how we connect it to a live NX session, with per-tool approval in front of every build tool, is covered on our Claude Code NX integration page.
Read Tools Before Build Tools: How an NX CAD MCP Server Is Structured
An NX CAD MCP server is a set of functions on NXOpen, each described by a name, typed parameters and a docstring, that an agent can discover and call. The design decision that matters most is not which library you use. It is splitting the tool set into a read tier the agent may call freely and a build tier that always passes through a human gate.
Our clear recommendation: ship the read tier first, on its own, and let designers live with it for two or three weeks before a single build tool goes live. Read tools are where the “the assistant finally sees my part” moment happens, and they carry no risk of corrupting a model.
What the Open-Source NX MCP Servers Do in September 2026
Two community servers define the state of the art for Siemens NX. The more mature one is NX_MCP (84 GitHub stars, 15 forks, MIT, retrieved 14 September 2026). It exposes 16 default tools plus 34 legacy tools behind an experimental flag. Journal execution is disabled by default. It was validated against NX 2506 (ugraf 2506.4021) on 21 August 2026.
The freshest signal is a draft pull request opened on 6 September 2026 that adds NX 2606 graphical-session hosting, durable mutation IDs with checkpoint rollback, geometry enumeration and roughly 189 capability-labelled tools, with a compact 13-tool agent profile on top. Whether it merges or not, it shows where community servers are heading: rollback and inspection are becoming first-class, which is exactly the production gap we describe on our service page.
The second server, a text-to-CAD demo built on NX 2206, creates primitives (box, sphere, cylinder), unions and differences, edge blends, and saves the .prt. It runs in batch mode only and its own README lists sketches, extrusions and holes as not yet implemented. Useful as a teaching example, not as a foundation.
Two data points from outside the NX world put this in perspective. Autodesk published two official Fusion MCP servers in April 2026, one local for modelling commands and one remote for design data, usable from Claude Desktop and Cursor. Siemens, by contrast, renamed NX to Designcenter with the 23 June 2026 release and shipped Designcenter Copilot, a documentation-grounded assistant, but no official MCP server for NX modelling operations exists that we are aware of. For NX teams that means the geometry layer is a build-it-or-partner decision in 2026, not a wait-for-the-vendor decision.
What We Built: A Chatbot With MCP Access to NX Geometry for an Asian Manufacturer
This year we delivered a chat assistant for an industrial manufacturer in Asia whose engineers wanted to talk to the part they had open, not to a help system. The assistant runs in a web panel next to NX and reaches the CAD session through an MCP server we built on the client’s NXOpen libraries. We do not publish client metrics for this project; what follows is the shape of the system, which is the part worth copying.
The read side went live first. Engineers ask “which features reference the wall-thickness expression?” or “what is the mass of body 3 in the current revision?” and the agent answers from the live model, with feature names and expression values it actually retrieved, not remembered. This alone changed how the tool was perceived: a colleague who knows the part, rather than a manual that knows NX.
The build side came four weeks later and is deliberately narrow. The agent can create sketch-based features, holes and patterns, and it can instantiate the client’s own template features, the boss-and-rib constructs their standards prescribe, with parameters filled from the conversation. Every build call shows a preview of what will change, waits for the engineer’s confirmation, sets an undo mark, and writes an audit entry. No free-form journal execution, no tool that takes screen coordinates.
The reasoning layer is a hosted frontier model, but the MCP server, the audit log and all part data stay on the client’s infrastructure. That separation is what got the project through the IT security review in one round instead of three. Our earlier work on a hybrid agentic assistant that executes and verifies inside NX followed the same pattern, and it is the pattern behind every action assistant in our NX AI assistants family.
A community server runs as a POC. What’s missing for production: role-scoped tool permissions, SSO/Entra, an audit log your CAD manager will sign, template awareness, on-premise operation. We close the gap from “runs on my machine” to “runs in the engineering department” — see how we build a gated NX MCP layer.
If you have generated NXOpen with an agent before, you know that the builder-and-section boilerplate is exactly where models hallucinate; we measured that in our NX Open code generation tests. Wrapping it once, by a human, into a typed tool means the agent never has to write it again. That is the real productivity argument for NX CAD MCP over “let the model write a journal”.
Why Confirmation Lives in the Protocol Now: MCP 2026-07-28
The July 2026 MCP specification changed the mechanics of “ask before you act”. The protocol core became stateless, the initialize handshake and session header were retired, and a new Multi Round-Trip Request pattern lets a server answer a tool call with resultType: "input_required", hand the client a question, and complete the call once the answer comes back. For an NX build tool that is the native way to say “this will add 24 holes to body 2; proceed?” without keeping a stream open.
Two more items from the same release matter to CAD managers. Enterprise-Managed Authorization was promoted to stable, so access to an NX MCP server can be governed through the identity provider you already run. And the extensions framework moved long-running tasks out of the core, which is where multi-minute operations like a full-assembly interference check belong.
In practice this means an NX CAD MCP server designed in 2026 should expose its gate through the protocol, not through a custom pop-up bolted onto the chat panel. We are moving our own servers to that model; the client project above still uses an application-level confirmation, and the migration is a contained change precisely because the tools were typed from the start.
What to Look For in an NX CAD MCP Setup
Whether you build in-house or with a partner like inteliScience, these six criteria separate a tool a CAD manager will sign from a weekend demo:
- Read tier ships alone first. Inspection tools with zero write access, in production, before any build tool exists.
- Engineering units in every signature. Millimetres, degrees, expression names. No screen coordinates, no “the selected face”.
- One undo step per write. Rollback must be precise, not “undo the last thing”.
- Template awareness. The agent instantiates your standard features rather than assembling geometry from primitives.
- Audit log the CAD manager owns. User, timestamp, tool, parameters, result, undo mark. Exportable.
- Model where you want it, data where it must stay. A hosted reasoning model is fine; the NXOpen bridge, the parts and the log stay on your infrastructure.
Number 4 is the one most demos skip and the one designers notice first. A boss built from a cylinder and a boolean is not the same as your standard boss template, and your works standard knows the difference.
Frequently Asked Questions
What is NX CAD MCP?
NX CAD MCP is the use of the Model Context Protocol to connect an AI agent to a running Siemens NX session through typed tools built on NXOpen. Read tools let the agent inspect features, expressions, bodies and PMI of the open part; build tools let it create or modify geometry after a human confirms the previewed change.
Can an AI agent really create geometry in Siemens NX through MCP?
Yes. Community servers such as NX_MCP already expose extrude, revolve, blend, hole and pattern operations, and we have shipped a client system in which the agent instantiates template features with parameters from the conversation. The constraint in production is not capability but governance: preview, confirmation, undo marks and logging around every write.
NX CAD MCP vs Designcenter Copilot: what is the difference?
Designcenter Copilot (the June 2026 release renamed NX to Designcenter) answers questions grounded in Siemens documentation and your current context. It does not execute modelling operations on your part. An NX CAD MCP server is the layer that acts: it reads the live model and, with your confirmation, changes it. Most teams end up wanting both.
How much does an NX CAD MCP server cost?
A read-only proof-of-value sprint on your own NXOpen libraries starts at €12,900 fixed price in our catalog and takes a few weeks. A production action layer with build tools, template awareness, gating, SSO and on-premise operation is a project from roughly €60,000 upward, depending on how many tools and templates it has to cover.
Is it safe to let an AI agent write to NX part files?
It is safe when the writes are narrow, typed and gated, and unsafe when the agent can execute arbitrary journals. We would advise against enabling any free-form code execution tool in an engineering department. Keep build tools to named operations with validated parameters, set an undo mark per call, and keep a log a human reviews.
Next Step
If your engineers are already describing parts to a chatbot by hand, the read tier is overdue and the build tier is closer than you think.