Is CAD Data Safe When AI Connects to Onshape?

Since releasing the FeatureScript MCP Server through Onshape Labs, the most common question we get isn’t about FeatureScript (Onshape’s native programming language), but some version of this:

“If I connect AI to Onshape, what happens to my design data?”

It’s a great question, and the answer will matter even more as AI becomes part of daily engineering workflows.

When connecting an AI model to other tools, most people picture a pipe that plugs together two systems, allowing data to freely flow between them. But that’s not the case. An MCP server is a toolkit Onshape provides to your AI model, not a data swap.

Let’s get into the actual mechanisms and how to keep your CAD data protected.

Understanding the MCP Setup with AI and CAD

Every MCP setup has three parts to make two systems work: the AI model, the MCP client, and the MCP server.

The AI model is the LLM, such as Claude, ChatGPT, or Gemini. This is a bring-your-own AI model setup, which means that your organization owns the account, the provider, and the contract. It should be vetted and approved by the team. Onshape doesn’t run it, host it, or see it.

The MCP client is where the work with the AI model happens. Engineers type requests in this workspace, in tools like Claude Desktop, Claude Code, an IDE, or a custom agent. The MCP client connects the AI model to the MCP server, passes requests between them, and displays approval prompts. Again, this is owned and managed by your organization, not Onshape.

The MCP server is Onshape’s toolkit. MCP (Model Context Protocol) is an open standard for how AI tools talk to applications, so the same server works with any compatible MCP client. The server offers a fixed list of tools.

For the FeatureScript MCP Server, the toolkit includes tools to search FeatureScript documentation, create, insert, test, and evaluate custom features. The AI model can use those tools and nothing else. For more background, read “What’s an MCP Server and How Does it Relate to CAD?”

Here’s a quick run-through of how these three pieces work together:

1. Through the MCP client, the engineer types a request, for example, “Create a custom feature that generates a roller chain sprocket,” for the AI model to reason.
2. The AI model calls on the MCP server for the relevant reference material to know how to perform an action.
3. The AI model writes the FeatureScript code and calls on a server tool to create a Feature Studio in Onshape. The MCP client asks for approval.
4. The code is evaluated in a sandbox workspace rather than against live design data. If it errors, the server will return the error message.
5. The AI model revises and retries, iterating in the sandbox until the custom feature behaves correctly.

The AI model asks for one thing, the MCP server does that one thing, and only the answer comes back. There’s no bulk sync.

Once the feature works, the AI model’s job is done. A finished custom feature runs like any other Onshape feature, can be easily shared with the wider organization, and used by the team without going back to the AI model. AI is involved only while the feature is being written, not every time it’s used.

Protecting Your CAD Data

To protect CAD data, understand that there are two important safeguards: Onshape’s built-in access controls and the AI you choose, whose provider’s terms apply.

Onshape’s Built-In Controls Apply

The MCP server works with the account’s permissions. The connection is authorized with an Onshape account through a browser step, the same flow as with any Onshape app. There is no separate permission model to configure and no elevated service account. Everything runs as the signed-in Onshape account, and actions are attributed to that account.

So a document that an account can’t open in Onshape stays closed to the MCP server and the AI model. Existing sharing rules are the access rules.

The MCP server can only do what its toolkit allows. The tool list is fixed, so if a task isn’t on the list, the AI model can’t do it. The AI model can’t browse an account. It can only ask the tools for specific things, within the account’s permissions.

Writes are limited and approved. An MCP server that writes does so only through its listed tools, under the same permission checks, and attributes each change to the signed-in account. The MCP client should ask the engineer to approve each write, so those prompts should stay on, and a first run belongs in a new or test document.

Choosing an AI Model

Keep in mind that the AI provider processes everything the AI model receives: the engineer’s prompts and every result an MCP server returns. Some of that can be valuable. Custom features built with the FeatureScript MCP Server, for instance, can capture a team’s engineering know-how. The FeatureScript MCP Server also stores a notes file in the Onshape account, and its contents go to the AI provider each time they’re read.

Across major providers, business and enterprise tiers typically exclude customer content from model training and specify how long data is retained under contract. Free and consumer tiers of the same products often don’t offer those protections.

Onshape can’t speak for a provider’s retention or training policies. Terms vary by provider and change over time, so a provider’s current data-use terms and data processing agreement deserve a careful read.

For proprietary designs, this decision matters most. It’s important to only connect to an Onshape MCP server with an AI tool that has been approved by your organization for that purpose.

Before Enabling an MCP Server

Six steps, in order of impact:

1. Choose a business or enterprise tier of the AI model, with terms that contractually exclude content from training and define retention. Avoid free and consumer tiers for proprietary work.
2. Have legal or procurement review the AI provider and its terms of service.
3. Review the MCP server’s tools and permissions to see what it can read, create, modify, or delete.
4. Choose which documents to share. Point the MCP client at new or test documents, and only at content the team is comfortable sending to the AI provider.
5. Decide who can subscribe using Onshape App Store controls.
6. Keep approval prompts on in the MCP client.

One more item to keep in mind is that every MCP server call counts toward the Onshape API allocation.

Connecting AI to Onshape with Confidence

AI belongs in the CAD workflow, and Onshape keeps it working within the rules teams already set. AI tools remain bound by the permissions of the Onshape account and can access only what that account can open.

The protection is built in. The only decision left is the AI provider, so choose one whose terms protect proprietary designs.

The FeatureScript MCP Server is available through Onshape Labs in the Onshape App Store. To get started, see “How to Get Started with the Onshape FeatureScript MCP Server.”