Canva MCP Server

Canva exposes its design platform to AI tools through a single hosted endpoint, https://mcp.canva.com/mcp. Authentication is OAuth — the endpoint's protected-resource metadata advertises scopes including design:content:read/write, asset:read/write, folder:read/write, brandtemplate:content:read/write, brandkit:read and comment:read/write — and Canva recommends Client ID Metadata Documents (CIMD) for client registration, with Dynamic Client Registration kept only for backward compatibility.

In practice the capabilities map to the everyday Canva workflow: generate a design from a text description, edit an existing design with natural-language commands, search your libraries and templates, manage uploaded assets and brand kits, add or read comments, and export finished work as PDF, PNG, JPG, PPTX or MP4. That last one matters for agent pipelines — 'make the deck, then export it as PPTX' completes end-to-end without opening Canva.

Client setup splits by transport support. Claude, Claude Code and Cursor connect to the URL natively (Claude also lists Canva in its connector directory). For stdio-only clients, the official docs publish an mcp-remote bridge config — command npx, args ["-y", "mcp-remote@latest", "https://mcp.canva.com/mcp"] — which proxies the remote server locally. Any Canva plan works for personal use; third-party products embedding the server need to register a redirect URI through Canva's waitlist form.

Trade-offs are the usual hosted ones: you can't self-host or extend the tool surface, per-tool rate limits apply (documented on Canva's 'Tools and rate limits' page), and complex multi-step edits still work better in the Canva editor itself. It's also a different product from the Canva Dev MCP server — the local @canva/cli one answers Apps-SDK documentation questions and doesn't touch your designs.