Draw diagrams privately with the local ZenUML MCP server
The local ZenUML MCP server draws diagrams for your AI assistant without sending the diagram source to ZenUML. It is the same tool as the ZenUML render server, but it runs as a small program on your own computer instead of on our servers.
Use it when your diagrams describe systems you would rather not send to a third party: internal architectures, unreleased APIs or customer flows.
Package: @zenuml/render-mcp on npm, Apache-2.0.
Local or remote?
| Local server (this page) | Render server | |
|---|---|---|
| Where your diagram source goes | Stays on your computer | Sent to zenuml.com/mcp, checked, not stored |
| Draws the diagram in the conversation | Claude Desktop | Claude and ChatGPT |
| Works in Claude Code | Yes, as text: no drawing | Yes, as text: no drawing |
| Works on claude.ai in the browser or on mobile | No: these only connect to remote servers | Yes |
| Needs | Node.js 20 or later | Nothing to install |
| Rate limit | None | 30 diagrams per minute |
Set it up
You need Node.js 20 or later. The first time your assistant starts the server, npx downloads it.
In Claude Desktop
Open Settings > Developer > Edit Config and add the server to claude_desktop_config.json:
{
"mcpServers": {
"zenuml-local": {
"command": "npx",
"args": ["-y", "@zenuml/render-mcp"]
}
}
}
Save the file and restart Claude Desktop. Claude asks for permission the first time it shows a diagram, then draws it in the conversation.
In Claude Code
claude mcp add zenuml-local -- npx -y @zenuml/render-mcp
Claude Code shows the result as text rather than a drawing. The server still checks the source and reports any errors, and the source still stays on your computer.
In other clients
Add a local (stdio) MCP server that runs npx -y @zenuml/render-mcp. Clients that support MCP Apps draw the diagram; others receive the result as text.
Supported diagrams
| Type | Use it for | Source |
|---|---|---|
| Sequence | Calls and messages between services or people | ZenUML |
| Mermaid | Flowcharts, state, class, entity-relationship, Gantt and other Mermaid diagrams | Mermaid |
| PlantUML | PlantUML diagrams | PlantUML, between @startuml and @enduml |
| Graph | Drawings you already have as draw.io XML | draw.io / diagrams.net XML |
| OpenAPI | REST API documentation | OpenAPI or Swagger, in YAML or JSON |
| AsyncAPI | Event-driven API documentation | AsyncAPI 2.x or 3.x, in YAML or JSON |
| Markdown | Notes and documents, with Mermaid blocks drawn as diagrams | Markdown; the assistant must say it is Markdown |
For every other type, ZenUML detects the language from the source when the assistant doesn't name it. If the source doesn't parse, the server returns each error with its line and column, and the assistant can fix the source and draw it again.
Use it
Ask in plain language, for example:
- "Draw a sequence diagram of how our payment service calls the fraud check."
- "Turn this AsyncAPI spec into documentation I can read."
- "Draw this draw.io file so I can see it here."
The server has one tool, render_diagram, and a short writing guide for each language as the resources zenuml://dsl-guide, mermaid://dsl-guide, plantuml://dsl-guide and openapi://dsl-guide.
Privacy
- Diagram source: checked on your computer and handed to your assistant to draw. It is never sent to ZenUML.
- What reaches ZenUML: requests for the drawing code, not your diagram. Your assistant downloads the diagram view's scripts, fonts and styles from
zenuml.com, and for draw.io diagrams the draw.io viewer and its shape libraries. These requests carry no diagram content, but they show your network address and which kind of diagram you drew. - Usage analytics: the view reports which kind of diagram it drew, whether drawing succeeded, when you zoom or switch to full screen, and (from version 0.1.3) that the diagram came from the local server. It reports no diagram content and no account or user identity. The program on your computer sends no analytics.
- PlantUML: drawn in the view by a built-in engine. No PlantUML server receives the source.
Your assistant still sends the conversation, including the diagram source, to its own AI model. Before you draw diagrams of confidential systems, check that this is allowed where you work.
For details, see the privacy policy.
Limits
- Source size: up to 500,000 characters per diagram.
- draw.io diagrams: the first drawing can take longer while the draw.io viewer downloads.