Bentley’s latest AI efforts are about connecting large language models to years of engineering know-how.
Bentley has sponsored this post.
AI is moving quickly into engineering software, and infrastructure companies are looking for more than what general-purpose tools can offer. While generative AI (like a GPT engine) is effective in understanding context, parsing unstructured information and exploring alternatives, it unfortunately cannot tackle complex engineering workloads with precision.
“The ability to calculate things exactly, leveraging well-known engines, modules and technology, is absolutely necessary because 90% accuracy is not good enough,” says Francois Valois, senior vice president of open application products at Bentley Systems. “You have to be 100% because lives depend on it.”
In this article, we’ll explore how Bentley is advancing the next step in infrastructure AI through the Model Context Protocol (MCP), an open standard that connects AI agents to validated tools, structured data and established engineering workflows.
What MCP does
In essence, MCP gives a large language model the ability to understand what another tool can do.
“It is like a business card for a piece of software,” explains Valois. “‘My name is STAAD, and I have the following abilities.’ When combined with a large language model or GPT types of technology, it opens up a whole world of possibilities for calculating deterministic things combined with generative input.”
STAAD has been around for more than 40 years as a structural engineering calculation engine, used across refineries, bridges, buildings and other major structures.
“People trust it, and it’s always providing the same results,” says Valois. “But on its own, it’s got limited scope.”
MCP enables STAAD to support larger enterprise workflows. An AI agent might pull information from different sources, identify the need for structural analysis, and turn to STAAD for a precise calculation. Based on criteria defined by a human engineer, it can run iterations, compare options and assess whether one result improves on the last. This all happens rapidly and at scale, with the engineer remaining in the loop and in control throughout.
“It’s really complementing existing workflows and augmenting human teams with agents as part of the team itself,” says Valois.
Bentley’s approach to MCP
Bentley has already published its STAAD MCP server to GitHub and the MCP Registry, and submitted it as a Claude Connector. Further, in September the MCP server was shipped natively with the STAAD.Pro update release. An MCP server has also been released for the complimentary product STAAD Foundation Advanced. It has also launched MCP servers for MicroStation and OpenSite+, and is working toward a fuller suite across its design products so that multiple servers and agents can be used together.
“I think this is where the real power of MCP unlocks,” says Valois. “It is not MCP alone, but MCP combined with the deterministic logic encapsulated into STAAD, MicroStation and others, enabling capabilities for broader consumption in agentic workflows.”
Bentley is building MCP into Copilot, the assistant embedded in its tools. The idea is to make Copilot understand the role of the person using the software. In OpenRoads, for example, a roadway designer could ask the system to create a certain type of corridor. Bentley uses retrieval-augmented generation (RAG) to ground the model in specific industry knowledge rather than relying on generic prompting.
Because Copilot is also an MCP client, it can connect to third-party servers to find answers, retrieve data or carry out other actions without the user leaving the application.
Bentley’s iTwin platform sits underneath this architecture, organizing fragmented engineering data—often trapped in files and disconnected systems—into something that is visible, standardized and semantically rich. It utilizes the Base Infrastructure Schema, a formalized way of describing infrastructure objects such as roads, bridges, pipes and pumps. Because iTwin is open source, data remains accessible throughout the entire lifecycle and across organizations.
“iTwin is absolutely a must-have because the results of AI are only as good as the data you feed it,” says Valois. “That becomes the basis for decision-making, AI automation and all kinds of different workflows.”
Data stewardship is another key consideration for Bentley.
“Your data is your data,” says Valois. “You have access to it at all times. You can retrieve it—we don’t monetize you getting out of our systems. We also don’t use your data to train our AI.”
Instead, Bentley draws on open-source data and synthetic data to train its neural networks—for example, for a tool that automates the placement of annotations on engineering drawings.
“We have a tool that allows you to place text on a drawing very fast in a way that’s aesthetically pleasing and readable,” says Valois. “With AI, you’re able to simulate beauty, which was not possible before.”
Bentley also signs explicit agreements with a limited set of customers to use their data. This opt-in model contrasts with common terms elsewhere in the industry that permit AI training by default unless a customer opts out.
MCP is the newest part of Bentley’s broader push around infrastructure AI, which spans OpenSite+ for site design, OpenUtilities Substation+ for substation design, and SYNCHRO+ for construction workflows and 4D planning. Valois says many firms are looking to bring their own agents to Bentley’s engineering logic.
“The only limit is the imagination of the humans involved in creating these workflows—which makes it fascinating and fast-changing.”
To learn more, visit Bentley MCP servers AI-powered engineering workflows
