The short version, for people deciding whether this fits. ODA MCP Servers launch in September 2026 — here is what they will do and how access works. For why it matters, see the announcement; for what they secure when your data can't leave, see the sovereignty piece.
The basics
What are ODA MCP Servers?
A connection between an AI agent and your engineering files. An MCP (Model Context Protocol) server lets any MCP-capable AI agent read and reason about a CAD or BIM model directly, asking it for a quantity, a property, a comparison, instead of being handed the raw file.
Does this replace my CAD software?
No. It is an access layer, not an application. It sits under your tools and lets an agent work with the model; you keep authoring where you author now. This is how AI reaches that data.
What will they do at launch?
Read and analyse models across all three formats: pull quantities, check properties, compare revisions, answer technical questions. They will also write back, and this varies by format: DWG goes furthest (create and edit entities, boolean operations, save to native .dwg), with lighter authoring in STEP and IFC. The strongest early results are in mechanical and drawing work.
Which file formats will they read?
DWG, STEP and IFC at launch. DGN, Revit and Navisworks are planned to follow.
Using it
Do I need to be a developer to use this?
To ask questions, no: you put them to the agent in plain language. To stand a server up and run it in-house, yes, self-hosting takes technical hands and hardware.
Does it work with Claude, ChatGPT, or my IDE?
Yes. It speaks the open Model Context Protocol, so any MCP-capable client works: Claude Desktop, Claude Code, VS Code, Cursor, ChatGPT-based agents, or your own. It is model-agnostic: run it against whichever language model you choose, local or hosted.
Can the agent actually see the model?
Yes. It can take a snapshot of the 3D view at any resolution and set the camera, angle, direction and render mode, before it looks. You can also mark up the model by hand, circle a part or drop a pin, and ask about exactly what you marked; your markup is baked into the image the agent sees.
What engineering numbers can it work out on its own?
Volume, surface area, centre of mass, mass and bounding box come built in, as single calls on the model's geometry. The math runs in fixed, tested code, so the numbers are reproducible rather than the model's best guess.
Can it work with several models at once?
Yes. Load STEP, IFC and DWG together, or run a server per format side by side, and the agent treats them as one surface. It can match a part across models, for example checking that the part defined in STEP sits where the IFC model says it should.
How is this different from just giving the file to ChatGPT?
A raw CAD file can't be read directly by a language model, and a text export is too large to parse reliably. The server works on the model through the same compiled engine that ships inside commercial CAD products, so the same model returns the same number every time and your tokens go to reasoning, not decoding.
Keeping your data in your control
Where does it run — does my data leave my network?
You host it yourself, on your own infrastructure. The agent works against your models inside your own network, and the data does not leave it. In local mode there is no network path out at all, the setup ITAR, TISAX and MDR environments call for.
Can the agent change my files, or only read them?
You decide. Every tool is marked for whether it only reads or can write, file access is whitelisted, and every call is logged, so there is a record of exactly what the agent looked at and touched.
Building on it
Can we build on it, or add our own formats?
Yes: it ships as an SDK as well as a product. You can stand up a server for your own format, write your own toolsets on top of the built-in ones, and drop in your own Python analysis scripts for the agent to call. A software vendor can embed it inside their own product.
Cost, timing, getting started
How much do they cost?
Royalty-free, and licensed per company rather than per seat: unlimited seats inside the company, with no per-query fee. They come with the Core Package+, included from the Sustaining membership tier up (Sustaining, Founding, Corporate); Commercial-tier members upgrade to reach it. The Extension formats are separate: Revit (BimRv) and Navisworks (BimNv) also need that Extension bought, while DWG, STEP, IFC and DGN need no add-on.
When do they launch?
September 2026.
Is there a demo I can look at?
At launch, the servers ship embedded inside ODA's reference viewers, Open IFC Viewer and Open STEP Viewer, where you can open a model and ask questions about what's on screen. Their source doubles as a working example for embedding the server in your own app.
How can I follow the launch?
Access comes with ODA membership at the Sustaining tier or above. Before launch, you can follow updates or talk to us.
New to the product? Start with the announcement and the piece on keeping your data in your network.
