What it is
Excalidraw MCP is the official MCP server from Excalidraw, the open-source whiteboard for hand-drawn style diagrams. It is an MCP App: instead of returning text, it renders an interactive canvas inside clients that support the MCP Apps extension, such as Claude, ChatGPT, VS Code and Goose. Claude's connector directory lists it as Anthropic verified, added in February 2026, with no sign-in required. The repo has about 5,560 GitHub stars.
The model gets five tools: read_me (a format reference it is told to call first), create_view, which streams elements onto the canvas with camera movement, export_to_excalidraw, which uploads an encrypted copy to excalidraw.com and returns a share link, and save_checkpoint and read_checkpoint. Each create_view call returns a checkpoint ID, so a later call can restore that scene, delete named elements and add new ones rather than redrawing from scratch.
The recommended setup is the hosted endpoint at https://mcp.excalidraw.com/mcp. You can also install a .mcpb bundle in Claude Desktop, build it and run it locally over stdio, or deploy your own copy to Vercel with no environment variables.
Install
https://mcp.excalidraw.com/mcpOur review
For chat users this is close to zero effort: add one URL, ask for a diagram, and an editable Excalidraw canvas appears in the conversation, filling in as the model draws. There is no account or API key, and you can expand the canvas to full screen and take over by hand. For explaining a system or sketching a flow during a call, that is the right weight of tool.
Editing is coarser than the checkpoint tools suggest. The model can restore a saved scene, delete elements by ID and add new ones, but it never sees a rendered image of the result, so it cannot tell when an arrow misses its box or two labels overlap; open issue #33 asks for arrow binding to be made easier for agents. The most-used community alternative, yctimlin/mcp_excalidraw (about 2,570 stars, with commits in October 2026), is built for coding agents instead: 26 tools for element-level edits, screenshots so the model can check its work, Mermaid import and .excalidraw files you can commit.
Diagram content goes to Excalidraw's servers. In the published code, a Vercel deployment with a Redis store keeps checkpoints for 30 days, and export_to_excalidraw encrypts the scene with an AES-GCM key that travels only in the share link's URL fragment, so excalidraw.com stores ciphertext. The connector listing points to the Excalidraw+ privacy policy without summarising it. The tools draw and fetch nothing from the web, so prompt-injection exposure is low, and a self-hosted copy keeps checkpoints on your own infrastructure.
Maintenance is the weak point. The last release, v0.3.2, shipped on 9 February 2026 and the last commit landed on 24 March 2026, while 37 issues and 27 pull requests sit open, including a dependency security fix submitted in May. The README still sends bug reports to the original author's repository, and GitHub detects no licence file although the README says MIT. It works today, but the feature set in October 2026 looks like the one adopters will have for some time.
Where it earns its place
A solutions engineer on a customer call, working in Claude
Sketch the customer's proposed architecture while they describe it
Prompt for the services and arrows, drag them into place on the full-screen canvas, then send the excalidraw.com link afterwards. It costs nothing; the trade-off is that the customer's architecture passes through Excalidraw's hosted server.
A backend developer documenting a service from VS Code
Turn a written request flow into a diagram for the README
create_view draws the first version and follow-up prompts restore the checkpoint and swap elements. Expect to fix overlapping labels by hand, and note the server writes no .excalidraw file into the repo; the community server does.
An engineering manager running a remote retrospective in ChatGPT
Draw the release process the team has just argued about
One prompt produces a flow the team can annotate live on the canvas. Free and instant, but the result lives in the chat and a share link rather than in a design tool the team already uses.
Before you commit
- Setup
- Add https://mcp.excalidraw.com/mcp as a custom connector with no sign-in, or install the .mcpb bundle in Claude Desktop; the first diagram follows the first prompt. Self-hosting is a fork-and-deploy on Vercel with no environment variables.
- Integrations
- Five tools: read_me, create_view, export_to_excalidraw, save_checkpoint and read_checkpoint. They write only to the canvas, the checkpoint store and an excalidraw.com share link; nothing touches your files or other services.
- Data and privacy
- Diagram elements go to the hosted server; the code caps input at 5 MB and keeps checkpoints in Redis for 30 days on Vercel when a store is configured. Exports are AES-GCM encrypted before upload. No retention statement or certification is published for the MCP endpoint.
- Cost at scale
- Free, with no plan or rate limit published for the hosted endpoint. The real cost is model output: create_view takes a full element array, so every revision of a large diagram is paid for in output tokens.
- Limits
- No image of the rendered scene for the model, no image insertion (issue #4) or PNG export (issue #31), and no canvas in clients without MCP Apps support.
What's good
- Anthropic-verified connector with no account or key
- Editable canvas renders inside the chat
- Checkpoints let the model amend a diagram instead of redrawing it
- Self-hostable on Vercel with no configuration
What's not
- The model cannot see the rendered result
- Canvas only renders in MCP Apps clients (issue #17: not visible in Cursor)
- No commits since March 2026; 27 pull requests unmerged
Best for
Chat users who want quick, editable diagrams in Claude or ChatGPT without any setup.
Alternatives
Excalidraw MCP: common questions
- Is Excalidraw MCP worth it?
- A one-URL way to put an editable sketch into a chat, but a coding agent that has to check and revise its own diagrams needs the community server instead. We rate it 3.8 out of 5.
- How much does Excalidraw MCP cost?
- Free hosted endpoint with no account; the code is MIT licensed according to the README.
- What are the downsides of Excalidraw MCP?
- The model cannot see the rendered result; canvas only renders in MCP Apps clients (issue #17: not visible in Cursor); no commits since March 2026; 27 pull requests unmerged.
- Who is Excalidraw MCP best for?
- Chat users who want quick, editable diagrams in Claude or ChatGPT without any setup.
- How much does Excalidraw MCP cost for a team or in production?
- Free, with no plan or rate limit published for the hosted endpoint. The real cost is model output: create_view takes a full element array, so every revision of a large diagram is paid for in output tokens.
- Is Excalidraw MCP safe for company data?
- Diagram elements go to the hosted server; the code caps input at 5 MB and keeps checkpoints in Redis for 30 days on Vercel when a store is configured. Exports are AES-GCM encrypted before upload. No retention statement or certification is published for the MCP endpoint.
- What are the best Excalidraw MCP alternatives?
- The closest alternatives we have reviewed are Figma MCP Server and Canva MCP Server.
Answers checked Oct 23, 2026.