Scribe-MCP demonstrates solid fundamentals: all 6 tools have clear verb-noun naming (vault_list, vault_read, vault_write, vault_delete, vault_search, vault_status), descriptions between 50 - 100 chars, and Zod-based input schemas with parameter descriptions. Tool annotations (readOnlyHint, destructiveHint) are present and correct. However, output schemas are not documented, responses are returned as JSON text blobs without formal schema declarations. Error handling is basic (try-catch with generic error messages) and lacks recovery guidance. Parameter descriptions are minimal (e.g., 'Note path, e.g. Projects/scribe/ideas.md' lacks format constraints). The vault_search tool accepts max_results with a default of 20, which is good, but no pagination cursor or total count is returned. Overall, the server is functional and well-named but lacks the polish and LLM-optimization of A-grade tools.

• Document output schemas for all tools. Define the structure of vault_list (array of {path, mtime, ctime, size}), vault_read (metadata + content), vault_search (array of {path, snippet, score}), and vault_status (document_count, folder_list, encryption_status). Include these in tool descriptions or as separate schema references.
• Enhance error messages with recovery guidance. Instead of 'Error: note not found: Projects/scribe/ideas.md', return 'Note not found: Projects/scribe/ideas.md. Try vault_list() to see available notes, or check the path spelling.'
• Add format constraints to path parameters. Specify: 'Path must be relative (no leading /), use forward slashes, and end with .md. Max 255 chars. No ../ traversal allowed.'
• Add an explicit inputSchema to vault_status with an empty object schema: {type: 'object', properties: {}, required: []}. This signals to clients that the tool accepts no parameters.
• Return pagination metadata in vault_search. Include total_count and has_more in the response so LLMs know whether to call again with an offset or cursor.