MCP server for sending ntfy.sh notifications when Claude needs user input
Two tools with basic schemas and descriptions. Both tools have well-defined Zod schemas and clear descriptions (98-104 chars each), which is within the ideal 10-1024 char range. However, several critical gaps reduce the score: (1) Parameters lack descriptions for enum values in request_user_confirmation (urgency enum lacks guidance on when each level is used). (2) No error handling guidance, failures return generic error messages without recovery suggestions (e.g., NTFY_TOPIC missing returns raw Error, not 'Set NTFY_TOPIC env var'). (3) Output schema is minimal text-only, not structured, responses don't return actionable data for downstream tools. (4) send_notification has optional parameters (title, priority, tags) with smart defaults, but these are not documented clearly in the description. (5) No security hardening: ntfy server URL and topic come from env vars (correct), but error messages may leak configuration details. (6) Tool composition: request_user_confirmation notes 'actual confirmation handling must be done separately', indicating incomplete design. No support for idempotency guarantees or retry safety.
Send a notification requesting user confirmation for an action. Note: This only sends the notification; actual confirmation handling must be done separately.
Send a notification via ntfy.sh to alert the user that input or action is needed
Enum parameter (urgency) lacks value descriptions. The enum ['low','normal','high'] is present but the description does not explain when to use each level (e.g., 'low: informational, normal: standard, high: critical'). This forces LLMs to guess mapping.
Error responses lack recovery guidance. When getConfig() throws 'NTFY_TOPIC environment variable is required', the error is not caught and wrapped with actionable recovery instructions. LLM receives raw Error, not 'Set NTFY_TOPIC=mytopic in your environment, then retry.'
Output schema is unstructured text-only. Both tools return {content: [{type: 'text', text: '...'}]}. If send_notification returns success, downstream tools cannot easily parse the topic or timestamp. Structured output with fields like {success: boolean, topic: string, timestamp: string} would enable tool chaining.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Incomplete tool design: request_user_confirmation explicitly documents 'actual confirmation handling must be done separately.' This signals the tool is only half-built and agents cannot complete a user-confirmation workflow in one interaction. Either implement bidirectional communication (e.g., polling) or add a receive_confirmation tool.
No idempotency guarantees. sendNtfyNotification does not support idempotent retries, if a retry succeeds, the user receives two notifications. Tools must be idempotent or document their non-idempotent nature with explicit side effects.
Optional parameters not documented for LLM guidance. title, priority, tags in send_notification are optional with smart defaults, but the description does not say 'If omitted, title defaults to "Claude Needs Input", priority to 3, tags to ["warning","claude"].' LLMs may not know these defaults exist.
No permission/scope declaration. Tools do not declare required permissions (e.g., 'send:notification'). This prevents least-privilege configuration and audit trails.