An MCP server gives an AI assistant actions it can take in a piece of software. An agent skill gives it instructions for doing a task well. A server is how the assistant reaches STAAD.Pro. A skill is what tells it which units the API expects and which order to build a model in.
They solve different problems, so they aren't really alternatives. On Tresca, 68 tools have both kinds of listing.
What an MCP server is
An MCP server is a program that exposes tools, data and prompts to an AI application through the Model Context Protocol. It runs as a process, locally or remotely, and does things: opens a project, reads results, changes a model.
What an agent skill is
The Agent Skills format describes a skill as "a folder containing a SKILL.md file". That file holds a name, a description and instructions for a task. A skill can also include scripts, reference material and templates.
The format was developed by Anthropic and released as an open standard. Agents load skills in stages to save memory: at start-up they read only each skill's name and description, and they read the full instructions only when a task matches.
For engineering, a skill is where working knowledge goes: the steps of a workflow, which checks to run, and the quirks of a particular program version.
Side by side
| MCP server | Agent skill | |
|---|---|---|
| What it provides | Actions and data in another system | Instructions, and sometimes scripts and references |
| What it is | A running program | A folder of files |
| Where it runs | On your computer or a remote service | Read by the agent; any bundled scripts run on your computer |
| Can it change your model? | Yes, if it has write tools | Not by itself; it needs a way to reach the software, such as a server |
| What you install | The server and its requirements | A folder in the place your agent looks for skills |
| Without the other | The agent has access but no guidance | The agent has guidance but may have no access |
Some skills do carry scripts that can act on files, so check what a skill includes before treating it as harmless text. The security questions apply to both.
How vendors are combining them
The clearest evidence that these work together is that software vendors now publish both.
Bentley and STAAD.Pro. The OpenSTAAD MCP Server runs OpenSTAAD Python against the model you have open. It comes with eleven STAAD.Pro skills covering geometry, loading, analysis, results and steel design. The core skill sets out the unit rules that the others depend on, which is exactly the sort of knowledge a server alone doesn't convey.
MathWorks and MATLAB. The MATLAB MCP Server starts or connects to MATLAB and runs code. The MATLAB Agentic Toolkit adds groups of skills for MATLAB and its toolboxes and installs alongside the server. MathWorks advises installing only a few skill groups, because agents trigger skills less reliably when many are loaded.
The OpenStudio team. OpenStudio-MCP provides tools to build and simulate building energy models. Its workflow skills cover simulation, retrofit studies and quality checks, and depend on the server's tools to do anything.
Equinor and NeqSim. The NeqSim MCP Server runs flashes, phase envelopes and process models. Separately, about 130 skills in the NeqSim repository guide an agent in writing and checking NeqSim code. In that case the skills also work without the server, because the agent can run NeqSim through Python or Java.
The community follows the same pattern. The plaxisMCP project pairs a PLAXIS 2D server with a skill recording the build order and known behaviour of PLAXIS 2D V20.
When you need which
Work through these questions for your own task:
Does the assistant need to read from or change the software itself? If so, you need a way in, usually a server. If you only want help writing input files or scripts that you run yourself, a skill may be enough.
Does the task have a procedure that a general AI model wouldn't know? Company standards, a specific sequence of analysis stages or the quirks of one program version belong in a skill.
Is there a server for your tool at all? Check the tool's page. Where there isn't one, a skill that helps write scripts or input files may be the practical route. Some tools are in that position, such as MCNP, which has a skill but no server.
Does your tool have an official server that only offers a generic "run this script" tool? Then the server provides access and little else, and a skill is often what makes it useful.
Who will maintain it? A skill is plain text that an engineer can read and correct. A server is software that has to be kept working as the engineering tool changes.
A note on how skills are distributed
Skills and servers often ship together, so one project can appear on Tresca as two listings. Some skills only work with the server they were written for, because they refer to its tool names. Each listing's limitations section says so where the project documents it.
Skills are also supported by more applications over time, but support isn't universal. Check that your AI application reads skills, and where it expects them. See which AI clients work for what listings document.
Whichever you use, AI-generated models and results still need checking by a qualified engineer.