MCP server for creating and editing draw.io diagrams through intent-level commands
Critical gaps in schema visibility and parameter documentation. Tool definitions are inferred from package.json and handler file names rather than explicitly visible in source. Input schemas exist but are severely underspecified: 'action' and 'operation' parameters are free-form strings with no enums, constraints, or validation guidance. Descriptions are present but generic (194-char baseline met for some tools, but lack actionable detail on when/why to use each). No output schemas documented. Error handling and recovery guidance absent. The server relies on natural-language command parsing (e.g., 'add TITLE', 'open PATH') which invites hallucination and misuse by LLMs.
Intent-level diagram operations: add/remove shapes, connect/disconnect edges, style elements, manage groups and layers, layout diagrams, and manage pages
Display the full verb reference and command syntax for draw.io operations
Query diagram state: list shapes, describe elements, inspect connections, search, compute statistics, view history, and compare against checkpoints
Session management for draw.io diagrams: create new diagrams, open existing files, save, manage checkpoints, and undo/redo operations
Free-form string parameters with no enums or constraints. 'action' in drawio_session accepts 'new TITLE', 'open PATH', 'save [as:PATH]', etc., but no enum or pattern validation. LLMs will hallucinate invalid syntax like 'new' without a title or 'open' without a path.
No output schemas documented. Tools return results but LLMs cannot plan downstream calls or extract required fields. E.g., drawio_query 'list' returns shapes but schema is invisible, LLM cannot know if it gets shape IDs, labels, or full objects.
Tool definitions inferred from file paths and package.json, not explicitly registered in source. Tool registration code not visible, cannot verify schema completeness or parameter descriptions in actual MCP registration.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | 2026-07-28+ | v2 |
No error handling or recovery guidance. If 'open PATH' fails (file not found, invalid format), no guidance on what to do next. LLM receives raw error with no actionable next step.
Parameter descriptions lack format/constraint details. 'action' description lists examples ('new TITLE', 'open PATH') but does not specify required vs optional parts, quoting rules, or error cases. LLMs cannot reliably construct valid commands.