An MCP server that lets AI agents draw hand-drawn Excalidraw diagrams. Real .excalidraw files on your machine, with history, rendering and a local library app.
Sketchdex demonstrates strong naming conventions (verb_noun pattern: list_scenes, describe_scene, read_elements, create_scene, add_elements, update_elements, delete_elements, replace_scene, import_svg, render_scene, delete_scene) and comprehensive parameter descriptions. All 14 tools have clear, actionable descriptions (avg ~150 chars, within 10-1024 baseline). Input schemas are present and typed for all tools. However, output schemas are not documented, callers cannot predict response structure. Error handling is minimal; no recovery guidance or error categorization. Tool descriptions excel at explaining WHEN to use each tool (e.g., 'describe_scene before you draw, render_scene after'), but lack explicit output field documentation. The INSTRUCTIONS block is excellent (pattern:tool-description compliance), but individual tool responses lack structured documentation.
Add elements to a scene. Shorthand for the common case: a shape gets a label (centred bound text), an arrow binds to named elements by ref, and generated geometry (polygon, star, arc, capsule, path) expands inline. Anything else passes straight through to Excalidraw.
Make a new scene, optionally with its starting elements.
Delete elements by id.
Delete a scene. It goes to trash, not the bin; recover it from the history sidebar.
What is on a scene, as structure rather than pixels: counts by type, overall bounds, and a rounded summary of each element with its id, position, size, text and bindings. Cheaper than read_scene and usually enough to plan an edit.
Draw an SVG as a single Excalidraw element. Use this for anything that has to LOOK LIKE something - a character, creature, object, vehicle, logo, illustration. Never assemble a figure from shapes; write it as one SVG and import it.
Output schemas not documented. Tools return complex structures (scenes with element arrays, rendered base64 images, paginated results) but callers cannot predict field names, types, or nesting. LLMs must infer response structure, risking parsing errors and wasted tokens.
No error handling guidance. Tools lack recovery hints (e.g., 'Scene not found. Call list_scenes to see available scenes.'). Errors are not categorized as retryable, user-fixable, or fatal. LLMs cannot determine next steps on failure.
Destructive operations (replace_scene, delete_scene) lack confirmation or dry-run support. Agents can irreversibly delete or overwrite scenes without a safety gate. Tool descriptions warn ('Dangerous: this is the only way to lose work') but no mechanism prevents accidental execution.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
Every scene in the vault with its id, name, element count and last-edited time. Start here.
Full, unabridged elements, filtered down to the ones you need. Use this rather than read_scene when editing a few things in a busy scene.
The whole scene exactly as stored, plus the path to the .excalidraw file on disk. Large scenes run to thousands of elements, so prefer describe_scene or read_elements unless you genuinely need everything. The path is there for edits no tool here covers - take a snapshot first if you use it.
Render a scene to a PNG. Returns a base64 image, or an error if no browser is found. Rendering is expensive; call it once per edit cycle, after describe_scene and the edits.
Replace the entire scene with a new element array. Dangerous: this is the only way to lose work. Take a snapshot first.
The manual. Topics: install, vault, app, rendering, tools, drawing, troubleshooting, or all. Start with install if you are setting this up.
Server version, vault path, and the MCP protocol version.
Move, resize, recolour, or edit the text of elements by id. Every field you pass replaces the old one; omit a field to leave it alone. Passing an empty object is a no-op.
ELEMENT_ARRAY schema uses additionalProperties:true with no type constraints. Agents can pass arbitrary fields; validation happens at runtime, not schema time. This invites hallucinated Excalidraw fields and silent failures.
Scene parameter accepts 'id, exact name, or unique part of name' but no validation rules documented. LLMs cannot predict which form to use; ambiguous matches may silently pick the wrong scene.