Pre-release · Coming Q3 2026

Core package

ODA MCP Servers

Connect any AI agent to the engineering data you already work with — whatever application created it.

Built on the ODA Platform — the engineering-data engine trusted by 1,200+ members for 25+ years.

  • For Software Vendors

    Add an AI layer to your product — without writing a line of protocol code.

    • Build on a proven engine.

      The same STEP / IFC / DWG core trusted across commercial CAD products, exposed via MCP.

    • Bring your own AI.

      Connect any MCP-capable agent — Claude, a local model via Ollama, whatever you run — so you’re never tied to one AI vendor.

    • Royalty-free.

      Part of the Core SDKs Package, on one fixed annual subscription.

  • For Manufacturers and Asset
    Owners

    Put AI to work on the engineering data you already own.

    • Keep it in your environment.

      Run the server on-premises with a local model, and your engineering data never has to leave your network — suited to air-gapped and IP-sensitive work.

    • Answers from the real model.

      Geometry, properties and quantities read straight from your files — never guessed.

    • Built to stay open.

      ODA is member-governed and can’t be acquired, so the access you rely on can’t be taken away.

Give AI the model, not the file

An AI can't work directly with a CAD or BIM file — the binary ones it can't read, and the text ones are too large and intricate to read efficiently. The raw file is the wrong thing to feed an agent.

ODA MCP Servers change that: the server sits between the agent and the file, so the agent works on the model — asking for the geometry, properties or structure it needs — not the raw file.

  • Deterministic
    answers

    Volume, mass and clearances run on the same compiled CAD math that ships in commercial products. Identical model in, identical number out.

  • No file too big

    The agent asks targeted questions;
    only the relevant pieces come back.

  • Lower token cost

    Parsing runs in compiled C++ — tokens go to reasoning, not decoding.

Your formats work together

A part lives in STEP, the building
in IFC, the detail in DWG.

Run a server for each and an agent can query them all in one session — same tools, same geometry schema, no custom pipeline.

  1. Run servers side by side — one per format, mounted together.
  2. One tool grammar, one geometry schema — shared across every server.
  3. Route each question to the right format — ask once, no glue code.

See it in action

AI agent builds a mountain road in DWG from a single prompt

DWG is now a target format for AI agents in the same way Python is a target language. This demo shows the full path: a prompt of a few lines asks for a mountain road through three named locations on real terrain, and the agent works with no human intervention from there.

From prompt to finished DWG

Terrain and routing in Python

It fetches and parses the terrain tile and runs the routing algorithm in Python on its own — then uses the ODA MCP Server to turn the result into CAD geometry.

DWG and PDF output

The output is a single DWG file with the full scene and a technical PDF report on the road, both produced by the agent. The DWG opens and edits as ordinary CAD data in OdaMfcApp or AutoCAD.

Plain language in.
Real answers out.

Numbers come from the same compiled CAD math that ships in commercial products — volume, area, surface area, mass, center of mass, clearances. Measured, not guessed.

  • “Which parts are missing a material spec?”

    7 of 142 — parts flagged

  • “Does this corridor meet the clearance code?”

    1.85 m — passes

  • “How much structural steel is on level 3?”

    12.4 tonnes — quantity takeoff

  • “List every fire door without a fire rating.”

    3 — doors

Deploy on your terms

Two deployment modes, one identical tool interface.

  • Local (STDIO)

    Runs as a local process with no network of its own, so it plugs straight into desktop agents and IDE extensions. Pair it with a local model and the whole pipeline stays on your machine — the fit for air-gapped and Intellectual properties-sensitive work.

  • Networked (HTTP/HTTPS)

    For multi-user, browser, or remote setups, or to embed the server inside your own product. Connections are secured with TLS.

Controls

The server only ever touches the folders you allow, and every tool is marked read-only or read/write, so you and the agent always know what can be read versus changed. Everything is logged.

Roadmap

  • At launch

    • Read — geometry and properties across STEP, IFC and DWG
    • Analyze — built-in volume, area, mass and center of mass
    • See — render the model, controllable camera, snapshots
    • Write the basics — add properties and create objects (incl. boolean ops on DWG), where the format supports it
    • Operate safely — file whitelisting, read/write tool annotations, logging

  • Next

    • More formats — DGN, Revit and Navisworks, then the rest of the ODA family, through the same agent-facing API.
    • Fuller authoring — geometry creation, entity transforms, boolean modeling and save-back to native files, beyond today's property and object writes.
    • Drawing automation — auto-generated orthographic, section and detail views, sheet layouts, and intelligent dimensioning and annotation.
    • Cross-format links — match objects across separate files and formats by different parameters (GUID, name, classification, or even geometric representation), so questions like 'is the STEP pump the same one in the IFC model?' can be answered.

Frequently asked questions

General

  • It's a program that exposes engineering model data — geometry, properties, structure — to AI agents through the Model Context Protocol (MCP), the open standard for connecting AI agents to tools and data. Instead of sending a whole CAD file to the AI, the agent asks the server for exactly what it needs. ODA MCP Servers do this directly for the DWG, STEP and IFC formats.

  • Yes. ODA MCP Servers work on the file format itself, not on a running copy of AutoCAD, Revit or any other program. Any DWG, STEP or IFC file works, regardless of which application created it.

  • Engineering files are large, binary and precision-critical. For binary formats it often isn't possible at all, and even when it is, the AI is guessing at the contents. The server reads the real model, so answers come from the actual geometry, not an approximation.

  • Any agent that supports the Model Context Protocol can connect — for example Claude Desktop, Claude Code and MCP-enabled IDE extensions — so it isn't tied to a single AI vendor. ODA's own viewers and demos currently run on Claude.

  • In STDIO mode your original files stay on your machine and are never uploaded. For remote or multi-user setups (HTTP/HTTPS) the connection is secured with TLS. File access is limited to folders you allow, and every action is marked read-only or read/write. Bear in mind the data the agent queries passes through the model's context, so choose the model accordingly for sensitive work.

  • ODA MCP Servers are part of ODA's Core SDKs Package, licensed royalty-free through ODA's annual subscription model; distribution terms scale with your membership tier. The product is pre-release, with launch planned for Q3 2026 — talk to ODA for current pricing.

  • ODA MCP Servers are coming in Q3 2026. You can talk to ODA now or sign up for launch updates; the reference viewers (Open IFC Viewer, Open STEP Viewer) show the servers in action.

For developers

  • Yes — ODA MCP Servers cover DWG/DXF, STEP (.stp/.step) and IFC (.ifc/.ifcxml) at launch. DGN (MicroStation), Revit and Navisworks are on the roadmap, through the same interface.

  • No. Files are read in their native format — no separate, lossy export step. Requested geometry comes back in a common JSON schema shared across all supported formats, so the same analysis tools work on any of them.

  • It runs in read-only or read/write mode. Read-only analyses without changing anything. In read/write it can author where the format allows — enriching models with new objects and properties, dimensioning. What's available is declared per format, so the agent only uses what MCP server provides.

  • Two transports, same tool interface. STDIO runs fully local with no network exposure and whitelist-restricted file access. HTTP/HTTPS is multi-client and session-aware, with TLS. In all cases the data the agent queries passes through the connected model's context — so for sensitive work, run a local model or make sure it is convenient for the case.

  • Yes. Several format servers can run side by side, sharing one tool grammar and one geometry schema, so the agent queries all of them through the same set of base tools and gets results in one uniform shape. That tool set keeps growing, and both ODA and members can add their own tools for their formats. Deeper cross-format links are on the roadmap.

  • Yes. ODA MCP Servers ship as a layered SDK. You subclass the parser for a format's core features and reuse the built-in tools, then add your own format-specific tools where you need them — you don't rebuild the server from scratch.

  • Yes. Drop in your own Python scripts and the agent can call them, the same way the built-in volume / area / center-of-mass tools work. You can also register reusable prompts and expose your application's data as resources the agent can use.

  • Two transports share one identical API. STDIO runs fully local with no network exposure, for IP-sensitive work. HTTP/HTTPS is multi-client and session-aware, with TLS — for browser-based, remote and embedded deployments.

Put AI to work on your CAD
and BIM data

ODA MCP Servers — coming Q3 2026.

Talk to us