The server defines 11 mail-specific tools with generally clear purposes, but exhibits significant issues in schema visibility, output documentation, and error handling guidance. All tool definitions appear in test files (e2e_test.go, live_test.go, mcp.go) with input schemas visible, but output schemas are completely absent from the source code provided. Parameter descriptions are concrete and domain-appropriate (e.g., 'notmuch query' syntax, MIME part numbers), but lack validation details (ranges, constraints). Tool names follow verb_noun convention ('count', 'search', 'show', 'thread', 'text', 'attach') but several are generic action words ('show', 'text') that lack object clarity. Error handling and recovery guidance are not evident in the code sample. The server is read-only except for 'refresh' (WRITE), which is appropriately risk-categorized but lacks confirmation/dry-run patterns. Schema completeness is weak: while input parameters have type and description, return types are undocumented.

• Document output schemas for all 11 tools. Specify return types (e.g., search returns array of {thread_id, subject, preview, date, senders}, with sample structure). This enables LLM chaining and prevents hallucinated field names.
• Rename generic tools: 'show' → 'show_message', 'text' → 'extract_message_text', 'thread' → 'get_message_thread', 'attach' → 'get_message_attachment'. Include the resource type in the name so LLMs pick correctly among 11 tools.
• Add validation constraints to numeric parameters: 'limit' should be 1 - 1000 (default 50 for search, 64K bytes for show); 'offset' should be ≥0; 'port' should be 1 - 65535. State these in descriptions or JSON Schema minimum/maximum.
• Enumerate valid account names in the 'account' parameter description. E.g., 'Limit to one account: alice@gmail, bob@proton. Omit for all accounts.' This prevents LLM typos and avoids silent fallback-to-all-accounts behavior.
• Add error recovery guidance to 'search', 'show', 'thread', 'text', and 'attach': 'If query is invalid, the error will list valid terms. If message ID not found, try search() first. If MIME part does not exist, call text() instead.'
• Consolidate the repeated 'query' parameter description into one canonical form and reference it via description comments, or define a reusable schema component to reduce redundancy across count, search, ids, files.
• Add a dry-run mode or confirmation step to 'refresh'. E.g., 'refresh' with account='*' and dry_run=true returns what would be synced without modifying state. Agents can then confirm before the real sync.

Inferred effective spec: <=2025-11-25.

• Relies on Dynamic Client Registration (deprecated) - use Client ID Metadata Documents (CIMD)