vega-mcp-server

Ask your AI assistant to draw a chart. Get back something interactive, embedded right in the conversation.

This is an MCP server for Vega-Lite and Vega, open-source grammars for building interactive visualizations from JSON. It works with any client that supports MCP Apps, including Claude, ChatGPT, Goose, and VS Code.

Browse the Vega-Lite and Vega example galleries for charts to remix.

Install

Claude Desktop (macOS / Windows)

Download the latest .mcpb file from the Releases page and double-click it. Claude will show an installation dialog and handle the rest.

Once installed, try asking: "Draw a bar chart comparing sales of [10, 20, 15, 25] for products A, B, C, D."

Other clients

See docs/INSTALL.md for build-from-source instructions and stdio / HTTP transport setup for Goose, VS Code, and ChatGPT.

Tools

show_vegalite_chart

Renders a Vega or Vega-Lite chart from a JSON spec. Pass data separately from the spec; it is injected at render time, keeping the spec clean and making validation straightforward.

Example

{
  "spec": {
    "$schema": "https://vega.github.io/schema/vega-lite/v6.json",
    "mark": "bar",
    "encoding": {
      "x": {"field": "category", "type": "nominal"},
      "y": {"field": "value", "type": "quantitative"}
    },
    "data": {"name": "table"}
  },
  "data": {
    "table": [
      {"category": "A", "value": 10},
      {"category": "B", "value": 20}
    ]
  }
}

Returns { success: boolean, errors?: string[] }. Validation runs the Vega-Lite compiler and checks that every named data source referenced in the spec has a corresponding entry in data.

Prefer Vega-Lite for most charts. Pass a raw Vega spec (with "$schema": "https://vega.github.io/schema/vega/v5.json") when you need capabilities Vega-Lite cannot express, such as custom layout transforms.

learn_vegalite_syntax

Returns a compact Vega-Lite v6 cheat sheet covering field types, mark types, encoding channels, and common transforms. Calling this before generating a spec helps the model stay within the supported grammar and avoid invalid field names. Kept intentionally brief; advanced features aren't documented here yet.

FAQ

Does it fetch external data or run arbitrary code?

No to both. Specs with remote data URLs are rejected before rendering, and the client-side loader blocks HTTP requests at runtime. All data must be passed through the tool call. The renderer uses Vega's sandboxed interpreter rather than eval, so arbitrary code execution is not possible.

Is it safe to deploy publicly?

This server was designed for local, single-user use. See docs/frequently-asked-questions.md for more.

Contributing

See CONTRIBUTING.md for development setup, pull request guidelines, and architecture decision records.

License

MIT