MCP in Practice: What Structural Engineers Need to Know Before Connecting AI to Their Software
clearspan
The previous post in this series explained what MCP is at a conceptual level: a standard interface that lets an AI model connect to external software, the same way a USB-C port lets any compatible device connect to any compatible cable. A structural analysis tool builds an MCP server once, and any MCP-compatible AI can use it.
That framing is useful for understanding why MCP exists. It does not answer the questions a practicing engineer needs answered before using it: How hard is it to set up? Who built this thing, and should I trust it? What happens to my client's project data? What does it cost?
This post covers those questions directly.
There Are Two Roles: Builder and User
"Using an MCP" means something different depending on where you sit.
Builders are the software developers who create MCP servers: the team at Bluebeam who built the Revu integration, the engineers at Clearspan who built the structural calculation connection, or a developer on GitHub who wired up a Tekla connector. Building an MCP server requires writing code, hosting infrastructure, and ongoing maintenance. This is not work for a structural engineer without a software background.
Users are the engineers and professionals who connect to existing MCP servers through Claude. No coding required. Setup means logging into a service and authorizing the connection, similar to connecting any cloud application. Once configured, the tools become available automatically at the start of your Claude session. You do not manually invoke them; Claude uses them when they are relevant to what you are asking.
What Connecting Actually Looks Like
Remote MCP servers run in the cloud, hosted by whoever built the integration. You connect the same way you connect to any web service: create an account, log in, authorize Claude to access the service. The server is already running on the provider's infrastructure. You are establishing a connection to it, not installing anything locally.
Bluebeam Max works in this fashion as well as the The Clearspan MCP. Once the connection is authorized, Claude can see the available tools and start using them in your sessions.
Who Can Publish an MCP, and What That Means for Trust
Anyone can publish an MCP server. There is no central authority that reviews or certifies them before they become available.
The result is a spectrum of trust:
Community MCPs are built and published by individuals or small teams, typically on GitHub. Many are well-constructed and actively maintained. Many are experimental or lightly tested. There is no independent review of whether the server handles your data responsibly, whether it has security vulnerabilities, or whether it will keep working after a software update.
Vendor MCPs are built by the company whose product you are connecting to. Bluebeam building an MCP for Revu, Clearspan building an MCP for its calculation platform. The quality, maintenance, and data practices follow the vendor's standard. You are trusting the vendor the same way you already trust them with your project data in their product.
Anthropic's reference implementations are maintained by Anthropic and cover general-purpose connections (web search, common developer tools) rather than AEC-specific software.
Anthropic maintains a directory at modelcontextprotocol.io that lists known MCP servers. Being listed there is informational, not a security or quality certification. Claude's integration store applies more formal review, closer to a curated app store. A server that appears there has received more scrutiny than a GitHub repository published last week.
The relevant question for practice is not whether an MCP is on a list. It is whether you would trust that software vendor with a copy of your client's project data, because that is what using their MCP involves.
What Happens to Your Data
Connecting an AI assistant to an MCP server puts your data in more than one company's hands. The AI provider processes the conversation itself, and the MCP provider receives whatever flows through its tools. When you create a calculation through the Clearspan MCP, that data is saved to your project record, which is the whole point. Each company's terms and privacy policy govern what it may do with the data, and those terms differ from provider to provider, and often between a consumer plan and a professional one. It is worth confirming how data is handled for the specific plan and tools you use.
For structural engineers, client confidentiality is a professional concern, and the practical question is simpler than the policy fine print: does the data you are sending contain information you need to protect? A beam analysis needs spans, loads, and materials, not the client's name or the project address. Some integrations let you enforce that automatically. The Clearspan MCP, for example, can strip client and project identifiers before anything reaches the model, so the AI runs the calc without ever knowing whose project it is.
What Sessions Cost
AI usage is billed by volume of text processed, measured in tokens. A token is roughly one word. The AI provider charges for tokens consumed during a session, both what you send and what comes back.
When you connect an MCP through a Claude subscription on the web or desktop, the tool calls draw on your existing plan's usage allowance rather than a separate per-token bill. Running the same kind of tools through an API-based product is billed directly by consumption instead. Either way, sessions that involve many tool calls (creating a project, running several calculations, generating a report) consume more than a simple question-and-answer exchange, though the cost is generally modest for focused tasks. Understanding which arrangement applies to you before running extended agentic sessions is worth doing.
A Short Checklist Before Using an MCP Integration
Before connecting Claude to any MCP server for project work, these questions are worth answering:
- Who built it, and do I trust them with project data? A community GitHub project and a vetted vendor integration carry different risk profiles.
- What does the provider's privacy policy say about data retention and use? Read it before connecting real project data.
- What data am I actually sending? Consider whether client-identifying information is necessary for the task.
- Who is paying for the AI usage, and how? Understand whether usage costs are included in a subscription or billed separately.
- Is this integration actively maintained? A server that has not been updated in a year may break with the next software release.
These are the same questions you should ask about any cloud software that touches project data. MCP makes it easier to connect AI to more services, which means the connections can multiply faster than the review process sometimes keeps up with.
This is the third post in Clearspan's series on AI for structural engineers. Earlier posts: LLMs for Structural Engineers and Skills, Tools, and MCPs. See it in practice: the Clearspan MCP for Claude.