@excaliwow/mcp

The Excaliwow Model Context Protocol server — lets an AI agent create, read, render, manage, and edit diagrams in your Excaliwow account through the same public REST API (/api/v1) the CLI uses. It runs over stdio, so any MCP client (Claude Desktop, Claude Code, etc.) can launch it with npx.

The public excaliwow/mcp repository contains distribution metadata and documentation for this package; the hosted application source is not published there.

Install

First mint a Personal Access Token at https://excaliwow.com/app/settings (Settings → Developer / API tokens) with read + write capabilities — enough for sixteen of the eighteen tools. Add delete only if you want the agent to trash and restore diagrams (see Security notes). Pass it as EXCALIWOW_TOKEN.

Claude Code (CLI)

One command. --scope local stores the server in your own settings, so the token never lands in a file you might commit. Export the token first so the literal PAT never lands in your shell history — --env NAME="$NAME" passes the value through without typing it a second time on the command line:

export EXCALIWOW_TOKEN=excw_pat_…
claude mcp add excaliwow --scope local \
  --env EXCALIWOW_TOKEN="$EXCALIWOW_TOKEN" \
  -- npx -y @excaliwow/mcp

Claude Desktop (and other JSON-config clients)

Add it to your client's config file (e.g. Claude Desktop's claude_desktop_config.json, or your client's MCP settings — see your client's MCP setup docs). The spec is unversioned, so npx -y always launches the latest release — you pick up new tools and fixes automatically, with nothing to bump:

{
  "mcpServers": {
    "excaliwow": {
      "command": "npx",
      "args": ["-y", "@excaliwow/mcp"],
      "env": {
        "EXCALIWOW_TOKEN": "excw_pat_…"
      }
    }
  }
}

The server reads EXCALIWOW_TOKEN per call from the environment (or, if you also use @excaliwow/cli and have run excaliwow auth login, that stored login) and never writes the token to disk itself.

Troubleshooting

"Not authenticated" / 401 / the agent's tool calls fail. The server starts even without a token (so it can list its tools), so a missing or invalid EXCALIWOW_TOKEN only surfaces when the agent first calls a tool. Starting with no token prints a one-line EXCALIWOW_TOKEN is not set warning to stderr. To check a token directly, run the health probe — it makes one authenticated read and prints a clear verdict (ok, 401 — token is invalid or expired, or could not reach <url> when the API itself is unreachable) and exits with a matching code (0 ok, 1 missing/invalid token, 2 unreachable):

npx -y @excaliwow/mcp --health

CLI flags

Tools

Eighteen tools, scoped to safe agent use:

read_diagram returns a summary + image, never the raw scene JSON, to keep context small. Pass includeGeometry: true to additionally get a compact, bounded { id, type, label, x, y, w, h } list (top-left x/y) so the agent can detect label/box collisions or misplaced nodes programmatically instead of eyeballing the PNG — it is derived from the scene, so it is present even when the render fails, and it is a small fixed-field summary, not the raw element dump. Pass includeScene: 'compact' instead for a bounded, element-addressable projection — a superset of the geometry fields that also carries style, text, and container/binding refs (≤17 fields/element) — the read half of the read_diagram → patch_elements surgical-edit loop. export_diagram returns the rendered bytes to save to a file — png as base64, svg as raw text — distinct from read_diagram, which returns an image block for a vision model to look at. An MCP server runs over stdio and cannot write to your repo, so a client with filesystem access (e.g. Claude Code) decodes and saves the bytes itself. Or skip the round-trip through the model and stream straight to disk with the CLI: excaliwow diagrams render <id> -o docs/architecture.png (or .svg) — also the fallback when a render is too large to return inline.

trash_diagram / restore_diagram are a reversible pair gated on the delete capability — registered always, they return a clean insufficient_scope error (changing nothing) unless the token carries delete, so a read + write token can't trash anything. Hard-delete/purge and making a diagram publicly shareable are deliberately not agent tools — those are irreversible, so a misled agent can't destroy or expose your diagram. Purge or publish from the dashboard or the CLI.

Two authoring paths

generate_diagram takes the high-level node/edge DSL and auto-lays it out — reach for it when you just want a quick flowchart. create_scene takes a raw, hand-authored Excalidraw scene, so the agent controls every element's position, size, color, stroke, fill, typography, and connections — the path for rich, polished, custom diagrams. You author elements tersely (id, type, geometry, text, colors) and the Excalidraw boilerplate is filled in for you; a fully-specified element (or a pasted .excalidraw scene) is passed through unchanged. To iterate on a rich diagram: get_scene → edit the elements → regenerate_diagram with the edited scene.

Discovery resources

The DSL grammar + a worked example are embedded in the generate_diagram tool description, with the full reference served as an MCP resource at excaliwow://dsl/reference. The rich-scene authoring primer + a worked example ride in the create_scene description, with the full guide (every element type, all styling props, bindings, groups, frames, palette, and layout/beauty heuristics) at excaliwow://scene/authoring.

Environment variables