An AI sketchpad with a memory and a feedback loop

What it does

It runs a live Excalidraw canvas alongside an MCP server, letting AI agents create, read, update, and delete individual diagram elements through 26 programmatic tools. Unlike the official Excalidraw MCP—which renders a one-shot diagram inline in chat—this keeps a persistent, WebSocket-synced canvas that multiple agents can manipulate concurrently. It also works without an MCP client by falling back to a REST API via an agent skill.

The interesting bit

The killer feature is the closed feedback loop: agents can call describe_scene for structured text or get_canvas_screenshot for an actual image of the canvas, then adjust specific elements rather than regenerating the whole diagram from scratch. It is essentially turning a generative toy into a visual REPL.

Key highlights

• Full element-level CRUD with layout helpers like align_elements, distribute_elements, and group/ungroup
• Canvas snapshots (snapshot_scene / restore_snapshot) and file I/O (export_scene / import_scene)
• create_from_mermaid for importing Mermaid diagrams
• read_diagram_guide embeds best-practice color palettes, sizing rules, and anti-patterns to improve AI output quality
• export_to_excalidraw_url encrypts and uploads scenes to excalidraw.com for sharing
• Multi-agent support: several agents can draw on the same canvas at once

Caveats

• The canvas server binds to localhost by default and has no built-in authentication; exposing it requires network-level access controls
• Requires two running processes (canvas server + MCP server), so it is more infrastructure than a single binary

Verdict

Worth a look if you want AI agents to iteratively refine complex diagrams rather than spamming prompt variations. Skip it if you just need a quick one-off sketch from a chat prompt—the official Excalidraw MCP is already there for that.