Editing is what makes it worth wiring up: nodes and edges go in and out of a stored board in a single atomic write, so a rejected operation leaves the file exactly as it was instead of half-changed. Markdown and SVG are the two export targets that exist today — PNG is written up as a proposal, not built.
An MCP server for JSON Canvas, the open infinite-canvas format Obsidian uses for its `.canvas` files. It writes canvases from node and edge objects, reads and edits stored ones, checks data against the JSON Canvas 1.0 specification, and exports a board to Markdown or SVG. Hosts that support the MCP Apps UI extension render a pan/zoom preview inline; text-only clients get the same canvas as structured output.
- A canvas written from nodes and optional edges into a date-prefixed `.canvas` file, answered with the path and the node and edge counts — `create_canvas`
- Stored boards read back as their nodes and edges, and the available files listed — `read_canvas`, `list_canvases`
- Nodes and edges added, updated and removed in one atomic write, where a failed operation leaves the file unchanged and removing a node cascades to its edges — `edit_canvas`
- Canvas data checked against the JSON Canvas 1.0 specification before it goes anywhere — `validate_canvas`
- A board exported as an edge-ordered Markdown outline or a standalone SVG image, returned with its MIME type — `export_canvas`
- Case-insensitive substring search across stored canvases, or scoped to one, returning the filename, id, field and a snippet per match — `search_canvases`
- Two resources to build against: the JSON Schema at `canvas://schema` and a two-node worked example at `canvas://examples/basic`
No account and no key. The published package is `mcp-server-jsoncanvas` 0.2.0 on PyPI, run with uvx over stdio; a container image ships at `ghcr.io/cam10001110101/obsidian-jsoncanvas`. `OUTPUT_PATH` is the directory canvases are written to and read from. Set `MCP_TRANSPORT` to `streamable-http` to serve over HTTP instead: it binds to 127.0.0.1 and carries no authentication, so anyone who can reach the port can read and write your canvas files — front it with an authenticating reverse proxy before exposing it, and set `MCP_CORS_ORIGINS` to the origins you trust rather than leaving the `*` default.
One command — docker run -i --rm -v canvas-data:/data mcp/jsoncanvas