Claude Code and Claude Desktop are the AI applications that engineering MCP servers and skills mention most often, followed by Codex and Cursor. ChatGPT appears far less, largely because most engineering servers run locally, next to the software they control.
The counts below come directly from Tresca's 412 listings and update whenever the site is rebuilt.
What the listings document
| AI client | Listings that document it |
|---|---|
| Claude Code | 237 |
| Claude Desktop | 190 |
| Codex | 125 |
| Cursor | 119 |
| VS Code | 75 |
| Windsurf | 23 |
| OpenCode | 22 |
| Cline | 20 |
| Gemini CLI | 17 |
| Claude | 16 |
| Google Antigravity | 13 |
| GitHub Copilot CLI | 12 |
Claude Code is named by 237 listings and Claude Desktop by 190. Cursor is named by 119 and ChatGPT by 11. A further 34 listings don't name any application in their documentation.
A listing often names several clients, so the numbers overlap and don't add up to the total.
What "documented" means
These counts record what each project's own documentation says. They don't mean Tresca tested the combination, or that the combination works today. They also understate compatibility: MCP is a standard, so a server may work with an application its author never mentions.
Read the numbers as a guide to which setups have instructions and which you'd be working out for yourself.
A listing's "AI clients documented" row is the place to check for a specific project. Where a project publishes install steps for an application, they appear on its listing page, with a link to where we copied them from.
Why local and remote matter most
The main thing that decides whether an application can use a server is how the server connects. MCP has two transports:
- Stdio, for local servers running on the same computer as the application.
- Streamable HTTP, for remote servers reached over a network.
Most engineering servers are local, because they have to reach software installed on your machine. Tresca lists 284 servers that support local connection. An application that can't start local servers can't use those directly.
How the main applications connect
This summarises each vendor's own documentation on the date at the top of this page.
Claude Desktop starts local servers from a configuration file, and the MCP documentation walks through it. See how to connect Claude.
Claude Code adds both local and remote servers from the command line. Its documentation covers claude mcp add, and lets you set a server for one project, for a whole team through a shared file, or for all your own projects.
Cursor supports stdio, SSE and Streamable HTTP. Cursor's documentation places project settings in .cursor/mcp.json and personal settings in ~/.cursor/mcp.json, and says it asks for approval before using MCP tools by default.
VS Code offers HTTP and "Command (stdio)" servers. Its documentation lists several configuration locations by scope and advises reviewing the publisher, command and requested access before starting a server.
ChatGPT connects to custom MCP servers through plugins, according to OpenAI's documentation, which describes remote servers reached by URL. That page doesn't describe local stdio servers. So a server that has to run next to your desktop software is likely to need extra work to reach ChatGPT, such as a tunnel. A few listings document one; the Horizun Civil 3D server, for example, names OpenAI's tunnel.
Features and plan requirements in these applications change often. Check the vendor's current documentation before you rely on this summary.
Servers tied to one assistant
Some vendor servers work only with the vendor's own assistant. The clearest example is Autodesk's AutoCAD and Civil 3D MCP Server, for which Autodesk lists only Autodesk Assistant. It's an MCP server in name and technology, but you can't connect Claude or Cursor to it. Autodesk's other servers differ, as our guide to Autodesk's MCP servers explains.
Vendor status doesn't tell you this on its own. Look at the clients row as well as the "Official" badge.
Skills are a separate question
Agent skills need an application that can read them, and not every MCP client does. Some listings also describe features specific to one application: the OpenStudio workflow skills, for example, use slash commands that are specific to Claude Code. For the difference between the two, see MCP servers vs agent skills.
How to choose
Start from the software you need to reach, not the AI application. Open the tool's page, see which listings exist, and check their clients, operating systems and how they connect. If your firm has already standardised on one AI application, filter for listings that document it.
Then work through the security questions. AI-generated models and results must be checked by a qualified engineer whichever application you use.