The Omnisend MCP server has 17 tools with basic naming conventions and schemas, but quality is inconsistent. Most tools follow verb_noun naming (list_contacts, create_product, delete_category), which is good. However, descriptions are generic and often under 100 characters, parameter documentation is minimal, and critical output schemas are not visible in the provided code. No error handling guidance is present in tool definitions. The server lacks input validation patterns, enum constraints for constrained parameters, and actionable error messages. Tool composition is reasonable (separate create/update/delete operations), but the server does not demonstrate the depth of schema documentation, error classification, or LLM-optimization patterns expected of production-grade tools. Most tools are CRUD wrappers with minimal intelligence.
Create a new product category
Create a new contact or update an existing contact
Create a new product in the catalog
Delete a product category
Delete a product from the catalog
Get detailed information about a specific category
Get detailed information about a specific contact
Get current brand information from Omnisend
Generic, minimal tool descriptions lack LLM-optimization context. Examples: 'Get a list of contacts from Omnisend' (42 chars), 'Create a new contact or update an existing contact' (52 chars). These provide no guidance on WHEN to use the tool, what it returns structurally, or dependency hints. Baseline expectation: 50-200 chars with 'what/when/why' framework.
Output schemas are not documented in visible source code. Tool definitions show input parameters but no explicit documentation of returned field types, structure, or required response fields. This forces LLMs to infer structure and plan downstream calls without confidence.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 47 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 46 | - | v1 |
Get detailed information about a specific product
Get a list of product categories
Get a list of contacts from Omnisend
Get a list of products from the catalog
Ping the MCP server to verify it is running
Replace an existing product with new information
Send a customer event to Omnisend
Update a category with new information
Update a contact with new information
'create_or_update_contact' violates single-responsibility principle. A tool name with 'and' signals multiple independent operations. This tool either creates OR updates based on input presence/format, making agent reasoning about which path executes unclear. Split into separate tools: create_contact (POST) and update_contact (PATCH).
No enum constraints visible for constrained parameters. Example: 'status' parameter in create_or_update_contact and update_contact accepts 'subscribed, unsubscribed, nonSubscribed' per description, but no JSON Schema enum is visible. This invites hallucinated values like 'active', 'inactive', 'pending'. Enums must be machine-parseable.
No error handling or recovery guidance in tool definitions. The code shows generic catch blocks that re-throw errors ('Error getting contacts list: ${error.message}'), but tool definitions do not indicate error classification (retryable vs user-fixable vs fatal) or corrective actions. LLMs receive no guidance on how to recover from failures.
Destructive operations (delete_product, delete_category) lack confirmation or dry-run patterns. No tool definition indicates that these are irreversible. Agents should have a confirm/rollback option before permanent deletion to prevent catastrophic errors.
Pagination parameters (limit, offset) lack explicit constraints. No documentation of minimum/maximum values, default behavior, or total record hints. The API response structure does not appear to return totalCount or next_cursor, making pagination logic opaque to LLMs.
Parameter descriptions are minimal (often matching just the input schema). Example: 'tags' parameter described as 'Tags to associate with the contact' without specifying max count, format constraints, or whether duplicates are allowed. These vague descriptions force LLM guessing.
Tool 'ping' has minimal semantic value for agent interaction. It returns no useful information and clutters the tool namespace. Consider removing it or folding connection verification into a status/health endpoint if needed. This follows the principle that tools should solve user problems, not internal diagnostics.