JSON Canvas MCP Server

A Model Context Protocol (MCP) server for working with JSON Canvas files — the open infinite-canvas format used by Obsidian. It lets an MCP client create, validate, read, and list .canvas files.

Built on the official mcp Python SDK (>=1.27), which negotiates the 2025-11-25 MCP protocol revision. Runs over stdio by default and optionally over the Streamable HTTP transport.

Hosts that support the MCP Apps UI extension render an interactive canvas viewer inline when you read or create a canvas — a pan/zoom, Obsidian-style preview of the nodes and edges. Text-only clients are unaffected and keep receiving the canvas as text/structured output.

Components

Tools

• create_canvas — Create a canvas from nodes (and optional edges) and write it to a date-prefixed .canvas file under OUTPUT_PATH.

  • Input: nodes (array of JSON Canvas node objects), filename (string, no extension), edges (optional array of edge objects).

  • Returns (structured): { path, node_count, edge_count }.

• validate_canvas — Validate canvas data against the JSON Canvas 1.0 specification.

  • Input: canvas (object with optional nodes and edges).

  • Returns (structured): { valid, error }.

• read_canvas — Read a stored .canvas file and return its nodes and edges.

  • Input: filename (string, with or without the .canvas extension).

  • Returns (structured): { nodes, edges } (also rendered by the canvas viewer; text fallback is the canvas JSON).

• list_canvases — List the .canvas files available in OUTPUT_PATH.

  • Returns: array of filenames.

• edit_canvas — Add, update, and/or remove nodes and edges on a stored canvas in one atomic write (a failed operation leaves the file unchanged).

  • Input: filename, plus optional add_nodes, update_nodes (partial, must include id), remove_node_ids (cascades connected edges), add_edges, update_edges, remove_edge_ids.

  • Returns (structured): { path, node_count, edge_count, canvas } — the updated canvas, so UI-capable hosts re-render it inline.

• export_canvas — Export a stored canvas to another format.

  • Input: filename, format (markdown | svg).

  • Returns (structured): { format, mime_type, content }. Markdown is an edge-ordered outline; SVG is a standalone vector image (node title lines only — plain SVG can't render Markdown).

• search_canvases — Case-insensitive substring search across stored canvases.

  • Input: query, optional filename to scope to one canvas.

  • Returns (structured): { matches: [{ filename, kind, id, field, snippet }] }.

create_canvas, read_canvas, and edit_canvas are linked to the canvas viewer via _meta.ui.resourceUri, so UI-capable hosts render the result inline.

Node objects use the JSON Canvas shape: id, type (text | file | link | group), x, y, width, height, optional color, plus type-specific fields (text, file/subpath, url, label/background/backgroundStyle). Edge objects use id, fromNode, toNode, and optional fromSide/toSide/fromEnd/toEnd/color/label.

Resources

• canvas://schema — JSON Schema for validating canvas files.

• canvas://examples/basic — A simple example canvas (two text nodes joined by an edge).

• ui://canvas/viewer.html — The interactive canvas viewer (MCP Apps UI), served with MIME type text/html;profile=mcp-app. Referenced by create_canvas and read_canvas.

Interactive canvas viewer (MCP Apps UI)

The viewer is a single self-contained HTML bundle built from the ui/ source with Vite and the official @modelcontextprotocol/ext-apps client. It renders nodes (with markdown, colors, and groups) and edges (sides, arrows, labels) in a pan/zoom view, themed via the host's CSS variables.

The built bundle is committed at jsoncanvas/_ui/viewer.html and ships in the package, so running the server needs only Python. Rebuild it after changing ui/:

make build-ui     # cd ui && npm install && npm run build  (requires Node.js)

To preview the renderer standalone (no MCP host), run cd ui && npm run dev and open /preview.html.

Related MCP server: MCP REST API Server

Usage with Claude Desktop

Docker (stdio)

docker build -t mcp/jsoncanvas .

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "jsoncanvas": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-v", "canvas-data:/data", "mcp/jsoncanvas"],
      "env": { "OUTPUT_PATH": "/data/output" }
    }
  }
}

uv (stdio)

{
  "mcpServers": {
    "jsoncanvas": {
      "command": "uv",
      "args": ["--directory", "/path/to/jsoncanvas", "run", "mcp-server-jsoncanvas"],
      "env": { "OUTPUT_PATH": "./output" }
    }
  }
}

Streamable HTTP transport

To serve over Streamable HTTP instead of stdio:

mcp-server-jsoncanvas --transport streamable-http --host 127.0.0.1 --port 8000

The MCP endpoint is then http://127.0.0.1:8000/mcp. The transport binds to localhost and, per the 2025-11-25 spec, validates the Origin header with DNS-rebinding protection enabled (localhost Origins only by default). To accept connections from outside the host (e.g. when running the container with HTTP), bind --host 0.0.0.0 and configure your allowed Origins accordingly.

Browser-based MCP hosts (the kind that render the canvas viewer) connect cross-origin and must read the mcp-session-id response header, so the Streamable HTTP transport serves permissive CORS headers. Restrict the allowed origins with MCP_CORS_ORIGINS (comma-separated; default *).

Security note. The HTTP transport is unauthenticated — anyone who can reach the port can read and write .canvas files under OUTPUT_PATH. The server is intended for local use; keep it bound to 127.0.0.1 (the default). DNS-rebinding/Origin protection is fixed to localhost Origins and Hosts at startup, so binding --host 0.0.0.0 exposes the port on the network but still rejects non-localhost Host/Origin headers — to safely expose it remotely, front it with an authenticating reverse proxy rather than publishing it directly, and set MCP_CORS_ORIGINS to the specific origins you trust (never *).

Configuration

Environment variables:

• OUTPUT_PATH — Directory where .canvas files are written/read (default ./output).

• MCP_TRANSPORT — stdio (default) or streamable-http.

• MCP_HOST / MCP_PORT — Host/port for the Streamable HTTP transport (default 127.0.0.1:8000).

• MCP_CORS_ORIGINS — Comma-separated allowed CORS origins for the HTTP transport (default *).

Development

# Install uv: https://docs.astral.sh/uv/getting-started/installation/
make setup        # uv venv && uv sync --extra dev
make build-ui     # rebuild the canvas viewer bundle (requires Node.js)
make test         # run the test suite
make lint         # ruff check + format check
make audit        # scan dependencies for known vulnerabilities (pip-audit)
make run          # run the server over stdio

Run the bundled library example:

make example      # writes example.canvas to OUTPUT_PATH (default ./output)

Library example

The jsoncanvas package can also be used directly:

from jsoncanvas import Canvas, TextNode, Edge

title = TextNode(id="title", x=100, y=100, width=400, height=100,
                 text="# Hello Canvas", color="#4285F4")
info = TextNode(id="info", x=600, y=100, width=300, height=100,
                text="More information here", color="2")  # preset color

canvas = Canvas()
canvas.add_node(title)
canvas.add_node(info)
canvas.add_edge(Edge(id="edge1", from_node="title", to_node="info",
                     from_side="right", to_side="left", label="Connection"))

import json
print(json.dumps(canvas.to_dict(), indent=2))

License

MIT. See LICENSE.