A low-level server implementation using the Model Context Protocol (MCP) with streamable HTTP support.
Single tool (create_note) with solid naming and reasonable parameter schema. Tool name follows verb_noun convention ('create_note'). Input schema is properly defined with type object, required properties, and field descriptions. Description is clear but terse. Tool definition is explicitly visible in ListToolsRequestSchema handler at src/index.ts line 95-109. However, output schema is NOT documented, the tool returns text but lacks a formal specification of response fields. Parameter descriptions are minimal (10-15 chars each). No error handling guidance in description. Tool is safe (WRITE) but no confirmation/dry-run pattern for data modification. Resources and prompts are well-structured with good URIs and descriptions, but the evaluation focuses on the tool.
Create a new note
Output schema not documented. Tool response structure (fields, types) is not formally specified. LLM cannot infer what fields are returned or plan downstream chaining.
Parameter descriptions are too brief (<20 chars each). 'Title of the note' and 'Text content of the note' lack guidance on format, length limits, or constraints. LLM has minimal context to fill values.
Tool description ('Create a new note') lacks WHEN guidance. Does not explain when to use this vs alternatives, what happens on duplicate titles, or prerequisites. Unclear if title must be unique.
No error handling guidance in tool description or response schema. LLM cannot infer what happens if title/content is empty, too long, or contains invalid chars. Validation happens server-side but is not communicated to LLM.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | C | 60 | <=2025-11-25 | v2 |
| 2026-03-09 | C | 60 | - | v1 |
No confirmation or dry-run pattern for a WRITE operation. Tool directly creates notes without user approval. Agents may accidentally spam notes without recourse.