Generate native, fully-editable draw.io diagrams from natural language — across all major AI coding assistants.
Server has 9 tools with complete input schemas and descriptions, but quality is uneven. Naming follows verb_noun convention (init_, add_, connect_, disconnect_, get_). Descriptions are present (avg ~120 chars) but lack actionable guidance on when/why to use tools. Schemas are well-structured with enums and type constraints, but output schemas are not documented. Error handling is absent, no recovery guidance or categorization. Three tools (add_container, add_node, connect) explicitly warn against misuse but lack positive guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Security: no secrets in params (good), but no audit trail or permission gates documented.
Only use for making incremental modifications or manual additions to an existing diagram. Do NOT use to construct new diagrams from scratch. Use compile_json_spec for new diagrams.
Only use for making incremental modifications or manual additions to an existing diagram. Do NOT use to construct new diagrams from scratch. Use compile_json_spec for new diagrams.
Only use for making incremental modifications or manual additions to an existing diagram. Do NOT use to construct new diagrams from scratch. Use compile_json_spec for new diagrams.
Connect a compute node to database and cache tiers with high-availability replication rules.
Bulk-connect all nodes in one tier to all nodes in another tier. Tiers are matched by node type or parent label.
Remove an edge between two nodes.
No output schemas documented. LLMs cannot predict what fields get_state or other tools return, forcing them to guess or make wasteful follow-up calls.
No error handling or recovery guidance. Tools lack categorization (retryable vs fatal) and actionable error messages. E.g., if connect() fails due to missing node IDs, the LLM has no guidance on what to do next.
Descriptions for add_container, add_node, connect warn against misuse but lack positive guidance on when/why to use them. Descriptions should explain prerequisites and expected outcomes, not just prohibitions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2026-07-28+ | v2 |
Get the current diagram state (cells, edges, metadata).
Initialize a new diagram. Call this before any other builder tool.
Connect compute nodes in primary and secondary AZs to a stateful data tier (RDS or ElastiCache) with full replication, local read/write, and cross-AZ write paths.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot distinguish safe reads (get_state) from destructive writes (add_node, connect). This prevents safe retry logic and planning.
Parameter descriptions lack format/constraint details. E.g., 'id' and 'parent_id' lack guidance on format (alphanumeric, max length, uniqueness). 'color' accepts hex but no validation hint. LLMs will hallucinate invalid values.