MCP: A New Standard for Connecting AI to AEC Software

Image source:  陈佳乐 /stock.adobe.com.

Model Context Protocol — MCP — is an open standard for connecting AI systems to external software tools and data sources. Originally released by AI company Anthropic in late 2024 and now governed as an open industry standard, it has attracted rapid adoption from major AEC software vendors who see it as a practical solution to one of the more stubborn problems in AI-assisted design work: getting AI to produce reliable, domain-specific results rather than plausible-sounding approximations.

The challenge is that while AI tools are helpful,  they can easily lead you down the wrong path. Ask a general-purpose AI assistant to calculate a structural load, and it will provide an answer. The answer will be confidently stated, well-formatted, but sometimes wrong. These systems are built on statistical prediction, not actual physics. For tasks where accuracy is expected — such as in building and infrastructure design — this is a fundamental problem.

MCP addresses this by connecting it to software that has the right answers. The protocol works by revealing a tool's capabilities to an AI system in a standardized way — defining what the tool can do and how to call it. When an AI encounters a task that falls within a connected tool's domain, it delegates the job to the tool. For example, a calculation goes to a certified analysis engine and the accurate result comes back. The AI handles the conversation, context-gathering, and orchestration. The specialized software handles the actual computation.

Because MCP is an open standard, it works across AI platforms. A server built to the MCP specification can be called by AI assistants such as Claude, Gemini, or GitHub Copilot, or any other that supports the protocol. That interoperability is a significant factor in the standard's appeal for enterprise AEC users, who may have existing AI investments or preferences that shouldn't be locked out by a single vendor's toolchain.

Here's where things stand, as of mid-2026. Many AEC software companies have now released MCP servers or announced they are in development, including Graphisoft.  Alternatively, there are third-party MCP servers for programs such as BricsCAD, Vectorworks, and McNeel & Associates’ Rhino and are available via the MCP market. DraftSight users can use the program’s API to embed their own AI assistant directly inside the program. Others do not have an official MCP, but are focusing on built-in AI assistance, such as TurboCAD Copilot. Editor's note: this technology is changing daily, so be sure to double check your program's current MCP development plan.

Autodesk, Inc.

Autodesk MCPs span both its AEC and design/manufacturing product lines, and it has two MCP integrations available. The Autodesk Product Help MCP Server gives AI agents access to up-to-date Autodesk documentation, improving accuracy when answering questions about product features and workflows. The Autodesk Fusion MCP Server enables AI systems to interact directly with Fusion designs — building features, editing geometry, and executing design tasks. We'll cover MCP for design and manufacturing workflows in a separate article.

A Revit MCP server is in technical preview as part of Revit 2027, released in April 2026. It ships as a separate add-on, downloadable from accounts.autodesk.com, and auto-configures with Claude Desktop and Cursor on installation. The current Tech Preview focuses on read-only access and includes seven tools covering model queries, element data retrieval, view navigation, and view export — giving AI assistants a structured way to read and understand a Revit model without touching or modifying it. Autodesk has said write capabilities — tools that would allow AI to create, modify, and manage Revit elements — are in development and planned for a dedicated write server, keeping read and write access cleanly separated.

You can learn more about Autodesk’s Revit MCP by attending the company’s upcoming webinar, “Making Complete Sense: MCPs & Autodesk Assistant in Revit 2027 — Everything Explained” scheduled for July 14, 2026.

Autodesk’s Revit MCP Server is in technical preview as part of Revit 2027. Image source Autodesk. Click image to enlarge.

Bentley Systems

Bentley has released MCP servers for two of its flagship products: STAAD.Pro, the structural analysis platform used on bridges, buildings, nuclear facilities, and large industrial projects; and MicroStation, its core CAD and infrastructure design environment.

The STAAD.Pro integration is designed to address the iterative nature of structural analysis. Engineers typically configure load cases, run calculations, review results, adjust parameters, and repeat — a process that limits how many design options can realistically be explored in the time available for a project. By connecting an AI agent to STAAD through MCP, that iteration can be automated. The AI can explore a wide range of design configurations, running each through STAAD's analysis engine, and surface the most promising results for engineer review. Bentley has demonstrated the approach on steel structure optimization, showing a 40% reduction in material weight through automated analysis runs.

In an interview with Francois Valois, VP of Civil Infrastructure at Bentley Systems, he explained, “If you can automate those [calculations] and be iterating based on the sort of space you’re looking to optimize, that makes it much more possible to iterate over that… Not only are we talking about doing things faster, but doing things better — augmenting the quality of the output.”

The MicroStation MCP server highlights a different use case: converting existing documentation into 3D geometry. Bentley has demonstrated an AI agent reconstructing a building in London using technical drawings, descriptive text, and 2D plans as input — producing a BIM (building information model) that engineers can continue to develop. “It’s helping re-engineer things that have been existing, maybe in non-digital form, to make them digital,” Valois explains, “and then be able to continue the work from there.”

Both servers run locally with no cloud data transmission, consistent with Bentley’s stated position that user data belongs to users and is not used to train Bentley’s AI systems. The servers are published to the open MCP registry and are compatible with multiple AI assistants, not just Claude.

MCP also fits into Bentley’s broader infrastructure data strategy. The company has developed a base infrastructure schema that formally defines objects — roads, pipes, structural members — making it possible to draw in data from a range of sources and expose it through AI-accessible workflows. Valois describes this as unlocking “dark data” or technical information trapped in CAD files and legacy documents that has historically been difficult to connect to live design processes.

A 3D structural analysis showing a meshed surface model in STAAD.Pro. Image source Bentley Systems. Click image to enlarge.

Bluebeam

Bluebeam’s MCP integration launched with Bluebeam Max in May 2026, a premium subscription tier for Revu. Bluebeam Max integrates Claude via MCP. Users can prompt the system using plain language to find information across project documents, create and update markups, build custom columns, and pull insights from markup data from within the Revu environment.

Additional features in Bluebeam Max include Smart Overlay, which identifies design changes and discrepancies across drawing sets; Smart Review, which flags scope gaps and missing information; and Magic Markups, which automates repetitive markup tasks. The tier also connects Bluebeam Studio Sessions directly to Revit, linking PDF markups to the corresponding locations in the BIM model.

While Bluebeam Max launches with Claude as its first AI integration, the company has indicated that its use of the open MCP standard is designed to allow future connections with other AI models, including Copilot, ChatGPT, Perplexity, and Gemini, as those platforms extend their desktop MCP support. Bluebeam is part of the Nemetschek Group, which also includes Graphisoft Archicad, Vectorworks, and Allplan. As of this writing these products have not yet announced official MCP integrations although some are in development and third-party MCP servers have emerged for some.

Bluebeam Max integrates Claude via MCP giving users a natural language path to automate Revu workflows. Image source Bluebeam. Click image to enlarge.

Trimble SketchUp

Trimble released its SketchUp MCP connector in April 2026, enabling users to create 3D geometry, such as building massing models, landscapes, furniture, by describing what they want in plain language, or by uploading sketches, photos, floor plans, and dimensions as reference material. The connector can generate models using both basic shapes and curved surfaces. Claude builds the geometry in a cloud SketchUp session, verifying dimensions iteratively, and tracks version history throughout so users can describe refinements or paste screenshots to point out specific elements that need adjustment. When a model is complete, the MCP creates a 2D preview thumbnail and it can then be downloaded as a SKP files compatible with SketchUp for Web, desktop, and iPad.

“The learning curve and time it takes for professionals to transfer a vision to a digital model has traditionally been the biggest barrier to 3D modeling,” said Chris Cronin, VP and general manager of architecture and design solutions at Trimble. The company framed the integration as part of a broader initiative to make 3D modeling accessible to users at any experience level.

Users need a Claude account and a Trimble ID to access the connector. Users receive up to 30 free saved SketchUp models; beyond that, a paid SketchUp subscription is required.

Trimble’s SketchUp MCP connector allows users to design and train Claude on core skills and workflows. Image source Trimble.

What Changes for AEC Teams, What Does Not

The practical effects of MCP will vary depending on the type of work and which tools you use, but a few consistent themes emerge across the vendor implementations.

The most immediate change is access. Engineers and designers will be able to perform tasks that previously required detailed software knowledge — knowing which command to use, which menu to navigate, which parameter to set — by describing what they need in natural language. The AI routes the request to the appropriate tool. For teams with many occasional users of complex software, this has the potential to reduce training overhead and make specialized analysis more accessible across a project team.

MCP can also make multi-tool workflows more feasible. Rather than manually moving data between applications, such as exporting from one tool, importing into another, reformatting, an AI agent with access to multiple MCP servers can in principle move information across tools automatically. Valois (Bentley) describes a scenario where an agent pulls data from enterprise systems, routes it through a structural analysis engine, and passes results into a CAD environment, all within a single session. “The real power is not calling one engine,” he says. “The real power is calling multiple of those engines [in an orchestrated workflow].”

MCP also carries implications for data that has historically been difficult to work with — technical content locked in older CAD files, PDF drawings, or text-based specifications. As AEC tools expose their capabilities through MCP, that information becomes searchable and usable in ways it wasn't before.

Running these workflows effectively also places new demands on hardware. AI agents coordinating multiple tools simultaneously are compute-intensive, and firms considering MCP-enabled workflows will need to evaluate whether their current workstations and local infrastructure can support them. For a closer look at what AI-ready hardware looks like for AEC environments, see Cadalyst’s recent article: AI-Ready Workstations for AEC Workflows. Plus we discuss how to get started with agentic AI and how to coordinate your workflows.

What MCP does not change is the engineer's role in making decisions. The tools connected via MCP are deterministic: they return accurate results based on valid inputs, but they cannot supply context or judgment that isn't there. “Without the human, none of that works,” says Valois. “You have to make sure people are still in control and fully involved in this — especially in the world of infrastructure.”

That point carries particular weight in AEC, where design decisions have long-term, physical consequences. The promise of MCP in this industry is not automation of engineering judgment, but reduction of the overhead between that judgment and its execution.

***