@deviva-draw/mcp

MCP server for Deviva Draw — lets AI agents (Claude Code/Desktop, Cursor, any MCP client) create, edit, query, and export whiteboard scenes with no browser involved. Everything runs headlessly on the same engine the draw.deviva.app canvas uses, so what an agent builds re-opens as a fully editable scene in the app.

Quick start

# Claude Code
claude mcp add deviva-draw -- npx -y @deviva-draw/mcp

# confine file access (open/save/export) to one directory — recommended
claude mcp add deviva-draw -- npx -y @deviva-draw/mcp --root ~/diagrams

Then ask the agent to draw:

Create a flowchart of a login process, check it looks right, and export it as login.svg

What agents can do

• Build scenes — shapes (rectangle/ellipse/diamond/sticky note/frame), lines, arrows, text, freehand ink; batch create/update/delete with auto-wrapped labels bound inside shapes.
• Generate diagrams — create_diagram (semantic nodes + edges) or create_diagram_from_mermaid, both with automatic layered layout and arrows that stay bound to their shapes.
• Self-verify — take_screenshot returns a PNG image block, so the agent can look at its own work and fix overlaps before finishing.
• Search, not dump — describe_scene, list_elements, and search_scene_content keep results token-cheap; full scene JSON only on explicit request.
• Export & round-trip — SVG and PNG exports embed the scene data, so exported files re-open editable in the web app; .devivadraw files (including multi-page documents) open and save losslessly.

With the desktop app

Point the server's --root at your drawings folder and open one of those .devivadraw files in the Deviva Draw desktop app: every agent save_scene live-reloads the open document within ~2s (camera preserved, unsaved local edits protected by a conflict bar — never merged or clobbered). Full walkthrough + troubleshooting: docs/desktop-agents.md.

PNG support (optional native canvas)

PNG export and screenshots use @napi-rs/canvas, an optionalDependency with prebuilt binaries for the common platforms. If it can't install (or you set DEVIVA_MCP_NO_CANVAS=1), the server runs in SVG-only mode: every other tool works, and the PNG tools explain the situation instead of crashing. Text measurement also uses the native canvas when present (browser-exact wrapping) and a calibrated approximation otherwise.

Security stance

• stdio only talks to files at paths the agent explicitly passes — no directory walking.
• --root <dir> (or DEVIVA_MCP_ROOT) rejects any path outside that directory.
• Image decoding accepts data: URIs only; the server never fetches remote URLs.
• Nothing leaves your machine.

Full setup (Claude Desktop, Cursor), the complete tool reference, and examples: docs/mcp.md.

License

MIT. Bundled hand-drawn font: Patrick Hand (SIL OFL 1.1).