MCP server giving an LLM read-only access to a personal Outlook.com mailbox via Microsoft Graph.
Three read-only tools with clear verb-noun naming (search_mail, list_recent_mail, get_message). Descriptions are present and adequate (100-150 chars), explaining WHAT each tool does and WHEN to use it. Input schemas are properly typed with descriptions for all parameters. Output schemas are documented in code but not formally declared in tool registration. Parameter defaults are sensible (max_results=15, unread_only=false). No destructive operations, so confirmation patterns not needed. Main gaps: output schema not formally documented for LLM consumption; no pagination guidance despite max_results cap; error handling is generic (RuntimeError with API status codes, not actionable recovery hints).
Fetch the full content of a single message by its id.
List the most recent messages in the inbox, newest first.
Search the Outlook mailbox for messages matching a free-text query. Searches across subject, body, sender and recipients (Graph $search). Returns message summaries; use get_message with an id to read full content.
Output schemas not formally documented. LLMs cannot infer what fields to expect from search_mail, list_recent_mail, or get_message. Code shows _fmt_message() returns {id, subject, from, received, isRead, hasAttachments, preview, webLink}, but this is not declared in tool metadata.
Error handling is generic and not actionable. RuntimeError with HTTP status codes (401, 4xx) does not guide the LLM on recovery. E.g., 401 error message says 'Try re-running outlook-mcp-auth', but the LLM cannot execute shell commands. Should return structured error with retry guidance or user-fixable hints.
No pagination metadata returned. search_mail and list_recent_mail cap results at 50 but do not return total_count or next_cursor. If results are truncated, the LLM cannot know whether to request more or if it has the full set.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Parameter 'query' in search_mail lacks format guidance. Description says 'Free-text search, e.g. movie tickets', but does not explain Graph $search syntax, special characters, or quoting rules. LLMs may pass malformed queries.
get_message returns raw body.content (HTML or plaintext) without format hint. LLMs cannot reliably parse HTML or decide whether to strip tags. Should convert to markdown or include bodyContentType in output for LLM awareness.