Canva MCP Server: Designs From Claude and ChatGPT

Canva MCP is Canva’s own Model Context Protocol server, and there are two of them. The one most people mean is the Canva AI Connector: a hosted server at https://mcp.canva.com/mcp that lets an assistant such as Claude or ChatGPT search your designs, generate new ones from a brief, edit text and images, add comments, resize, and export to PDF, PNG, PPTX or MP4. You sign in with your own Canva account, and the assistant can only do what you could do yourself. The second is the Canva Dev MCP server, which runs on your machine and gives a coding agent Canva’s developer documentation while you build a Canva app. It does not touch your designs at all.

Two servers, and what each is called

Canva’s naming shifts depending on where you read it. The help center calls the user-facing product the Canva AI Connector. The developer site calls the same thing Canva for AI assistants, one of three surfaces of the Canva Developers SDK, which now brings the Apps SDK, the Connect APIs and MCP together. Canva’s overview of Canva MCP↗ (opens in a new tab) says it is a remote server exposing design creation and editing, asset and brand management, library search, export and commenting as tools. Neither server carries a beta or preview label in Canva’s docs today, though the developer pages note that Canva is still improving the server.

• Canva AI Connector: remote, OAuth sign-in, acts on your designs. For anyone with a Canva account, including Canva Free and Canva Education.
• Canva Dev MCP server: local, started with npx, answers questions about the Canva Developers SDK and the App UI Kit. For developers building Canva apps.

Setup in Claude

In Claude on the web or the desktop app, open Settings, choose Connectors, find Canva and select it, then follow the prompts to sign in to Canva and approve access. Start a new chat, open the settings menu in the chat box, and make sure Canva is toggled on under Connectors. The same toggle turns it off for one conversation. To remove it for good, go back to Settings, choose Connectors, open the menu next to Canva and select Disconnect. Canva’s Claude setup guide↗ (opens in a new tab) walks through each screen.

Setup in ChatGPT

ChatGPT now calls its connected apps plugins. Open Settings, choose Plugins, find Canva and follow the prompts to connect your account. You can also just ask for a design and mention Canva; ChatGPT offers to connect it the first time. On ChatGPT Business and Enterprise, a workspace admin has to allow apps and connectors before anyone can add Canva. In ChatGPT, generated designs appear as previews first, and only the one you pick is saved to your Canva account. Canva also has guides for Gemini and Microsoft Copilot on the same help pages.

Coding clients such as Claude Code

Canva’s user guides cover the chat apps. For other clients, its developer docs give the server address and a condition: Canva has to recognize the client’s OAuth identity first. A product can register an app in Canva’s Developer Portal, or, if it identifies itself with a Client ID Metadata Document, ask to be added to Canva’s allowlist; Canva says many popular AI tools already publish one and work out of the box. If your client is one of them, adding the server is one line, and the first tool call opens a browser to sign in:

claude mcp add --transport http canva https://mcp.canva.com/mcp

If sign-in is refused, the client is probably not one Canva recognizes, and no setting on your side changes that. The general flow is explained in MCP OAuth, explained.

The Canva Dev MCP server, for app builders

The Dev MCP server page↗ (opens in a new tab) lists Cursor, Claude Desktop, Claude Code and VS Code, and needs git, npm and Node.js 24 or later. It runs locally, fetches documentation from canva.dev, and Canva says it sends your code and prompts nowhere except your own AI tool. To make the agent use it, name the thing you want, such as the App UI Kit or the Canva Developers SDK, in the prompt.

claude mcp add canva-dev -- npx -y @canva/cli@latest mcp

{
  "mcpServers": {
    "canva-dev": {
      "command": "npx",
      "args": ["-y", "@canva/cli@latest", "mcp"]
    }
  }
}

What it can create and edit

The connector’s tools fall into a few groups. Their names use hyphens, so you will see calls such as search-designs in your client’s approval prompts.

• Find and read: search-designs, get-design, get-design-pages, get-design-content for the text on a design, get-presenter-notes, and folder and asset listings.
• Generate: generate-design turns a brief into several candidates, and create-design-from-candidate saves the one you choose as a real, editable design. Canva tells integrators to show you the candidates rather than pick one for you.
• Edit: changes happen inside a transaction. start-editing-transaction opens it, perform-editing-operations replaces text or swaps images, and nothing is saved until commit-editing-transaction; cancel-editing-transaction throws the draft away. An image has to be uploaded with upload-asset-from-url before it can go into a design.
• Copy, resize and import: copy-design, resize-design for a new size or channel, and import-design-from-url to bring a file in as an editable design.
• Comment: comment-on-design, reply-to-comment, list-comments and list-replies.
• Export: export-design produces PDF, PNG, JPG, PPTX, MP4 and other formats. The download links are signed and expire, so save the file rather than the link.

Canva asks every integration to end a design task with a link that opens the design in the Canva editor, and its design edit handoff guide↗ (opens in a new tab) is blunt about why: an export is a static file, and the person still needs to review and adjust the result. Treat that link as the real output. Prompts that work well name the format and the audience, for example: “Search my designs for the spring flyer, change the date line to read Saturday, April 18, and give me the edit link before exporting a PDF.”

Brand kits and brand templates

Brand features need a paid Canva plan. list-brand-kits reads your brand kits, search-brand-templates finds templates shared with you, and create-design-from-brand-template starts a design from one. Autofill (autofill-design, with get-brand-template-dataset to read the fields) fills a template’s fields from your data, which suits batches of near-identical designs such as event cards or product tiles. Canva’s pages do not agree on which paid plans include autofill: the tools table says Pro and above, while the ChatGPT guide lists it as Enterprise only. Check your own plan before you build a workflow around it. Only brand templates that have been shared with you show up.

Limits

• Per-tool rate limits. Canva’s tools and rate limits table↗ (opens in a new tab) sets a per-minute cap on each tool: 100 requests a minute for most reads, and 20 for generating, copying, resizing, exporting and each step of an edit.
• Plan-gated tools. Resizing, autofill, brand kits and brand templates are Pro and above. Export works on every plan, but quality depends on it, and a design with premium elements can fail to export with license_required.
• Slow generation. generate-design can take a while; Canva suggests a 60-second timeout for clients that give up early.
• Credits and cooldowns. Some actions use Canva AI credits and can hit a limit or cooldown, on top of your assistant’s own message limits.
• Not everything in Canva is a tool. The list is what the server exposes today, and the design still gets finished in the Canva editor.

Permissions and admin controls

Every user signs in individually; Canva does not offer a shared service account for the MCP server. What the assistant can do matches your own access, so it can edit only designs and assets you can edit. On Canva Teams, Business, Nonprofit, Education and Enterprise accounts, the connector counts as a third-party integration, and an admin can turn it off under Admin Dashboard, Controls and Permissions, Third-party integrations. When they do, nobody on the team can connect an assistant and existing connections stop working.

Two habits keep it safe. Keep approval prompts on for tools that write, comment or export, because they act as you on designs other people may be using. And treat design text and comments as input written by whoever can edit the design; anything an agent reads can carry instructions, which is indirect prompt injection. The wider checklist is in MCP security best practices.

Canva, Figma or Miro

The three design-side servers answer different jobs. Canva’s makes finished marketing and presentation assets: social posts, flyers, decks, videos. The Figma MCP server hands a coding agent frames, components and variables to turn into product UI. The Miro MCP server reads and writes whiteboards, which suits planning, flows and diagrams.

Turning design requests and feedback into tasks

Design work arrives as requests and leaves as comments, and both get lost in chat. With fenbs connected as a second MCP server at https://fenbs.ai/api/mcp, an assistant can read the open comments on a design with list-comments and file each change someone asked for as a task in To Do: an enhancement or a bug, with the Canva edit link in the note and priority left for a person to set from 1 (most urgent) to 10. It should search the board first so it does not file the same request twice. If brand voice or color rules must hold for every design, record them as a rule on the Decisions and rules page, which every connected assistant reads before it starts.

Be clear about the gaps. fenbs has no due dates and no assignee field, so a launch date or a designer’s name goes into the note as text. And fenbs does not store designs; the design stays in Canva, and the task holds the link to it.