An MCP server is a small program that lets an AI assistant use another piece of software. A server for a structural analysis package, for example, might give the assistant actions such as "list the load cases" or "read the member forces". The assistant decides when to use those actions, and the server carries them out.
MCP stands for Model Context Protocol. The project's documentation calls it "an open-source standard for connecting AI applications to external systems" and compares it to a USB-C port: one standard connection in place of a different cable for every device.
For engineers, the practical result is that an assistant such as Claude can work with the model open on your screen instead of only talking about it. Tresca lists 323 servers and plugins and 89 skills across 120 engineering tools.
The three parts
Every MCP setup has the same three parts, which the MCP architecture page defines:
- The host is the AI application you use, such as Claude Desktop, Claude Code or VS Code.
- The client is the component inside that application that keeps a connection to one server. You rarely deal with it directly, and people often say "client" to mean the whole application, as Tresca does on its listing pages.
- The server is the program that offers actions and information.
A server offers three kinds of thing: tools, which are actions the assistant can run; resources, which are data it can read; and prompts, which are reusable templates. Most engineering servers are mainly collections of tools.
An MCP server doesn't contain any engineering knowledge of its own. It exposes what the software allows and leaves the AI model to decide what to do with it. That's why agent skills exist as a separate idea.
How a server reaches engineering software
An MCP server can only do what the engineering tool lets programs do. Each tool page on Tresca describes that tool's automation interfaces for this reason. Servers reach the software in one of four main ways, and Tresca records which one each listing uses:
| How it connects | What it means | Example |
|---|---|---|
| Drives a running copy | The server talks to the program open on your computer through its API | OpenSTAAD MCP Server needs STAAD.Pro running with a model open |
| Reads and writes files | The server works on files directly, and the program may not be needed | DWG MCP Server reads DWG files without AutoCAD |
| Vendor's cloud API | The server works through the vendor's online service | Autodesk Fusion Data MCP Server reaches hubs, projects and items in Autodesk's cloud |
| Documentation only | The server answers questions and doesn't operate the software | Autodesk Product Help MCP Server searches Autodesk's help content |
Of the servers listed here, 251 drive a running copy of the tool. That has a practical consequence: the software usually has to be installed and licensed on the same machine, and 149 servers run on Windows only, because much engineering software does.
Local and remote servers
MCP has two ways of carrying messages, and the difference matters when you assess a server.
A local server runs on your own computer, started by your AI application. It usually communicates through what the specification calls the stdio transport, which is direct communication between two programs on one machine. Most engineering servers work this way, because they need to reach software installed on that machine.
A remote server runs elsewhere and is reached over the internet using the Streamable HTTP transport. Vendor-hosted services work this way.
The documentation on connecting local servers makes one point worth remembering: a local server runs with your user account's permissions, so it can do anything you could do yourself. See the security questions to ask before installing one.
What using one looks like
Once a server is connected, you ask for something in plain language. The AI application offers the server's tools to the model, and the model chooses which to call. Claude Desktop, for example, asks for your approval before running a tool in the documentation's worked example. How approval works varies by application and setting, so check yours.
Some servers can only read. Others can change your model, and some offer a choice. Autodesk's Revit Public MCP Server, for example, comes in Read, Write and Experimental configurations, and only the first leaves the model unchanged. Each Tresca listing states which applies.
Who makes them
Servers come from two main sources. Software vendors have started publishing their own, and our tracker of official releases currently shows 37 servers and 10 skills from vendors and tools' own developers. Many of these are still labelled beta, preview or pre-1.0.
Everything else comes from the community: individual engineers, research groups and companies. Quality varies, so read each project's documentation and check when it was last updated. Some tools have no integration at all yet, and those are listed on our wanted page.
What an MCP server doesn't do
An MCP server doesn't make an AI model correct. It gives the model access to the software, and the model can still choose unsuitable parameters, misread a result or apply the wrong assumption. A server that runs a calculation successfully tells you only that the calculation ran.
Treat anything produced this way as a draft for an engineer to check, in the same way you would treat work from someone new to the project.
Tresca describes what each project documents. We don't install or test these servers, so a listing is not a recommendation.