Local stdio MCP mail, calendar, and weekly-report assistant for on-premises Microsoft Exchange via EWS + NTLM.
This server has significant quality gaps despite reasonable naming conventions. Of 25 tools listed, 7 (tools 17-23, save_mail_draft, edit_mail_draft, continue_action, read_calendar, find_meeting_times, save_meeting, send_meeting_invitation) have ONLY placeholder descriptions ('Production tool profile implementation for...') with no actual content, these score 0 on description. This alone caps their per-tool scores at 0-30. Tools 1-16 have better descriptions in Chinese/English mix, but schemas are NOT explicitly visible in the provided source code, only function signatures with type hints are shown. Per the rubric's HARD SCORING RULE: 'If you cannot see the actual tool definition in the source (only inferred), cap that tool's overall at 50.' Most schema documentation is inferred from Python function signatures, not from explicit JSON Schema declarations. Tools 17-23 are defined in tool_profiles.py but with zero substantive description. Parameter descriptions vary: tools 1-16 have inline documentation in Chinese, but tools 17-23 lack any param-level documentation. Error handling is not visible in the provided code excerpt. Output schemas are not documented anywhere in the provided source. The README link suggests this is for on-premises Exchange via EWS/NTLM, a domain-specific tool, but the quality presentation is inconsistent and incomplete.
把允许目录中的本地文件添加到现有草稿,绝不发送。item_id/draft_ref 二选一;优先使用 draft_ref。
解析收件人并创建邮件草稿,绝不发送。查询只使用姓名拼音或完整邮箱。存在无法自动消歧的收件人时,返回 resume_token,由 continue_action 接收用户选择后继续创建草稿。
Production tool profile implementation for continuing multi-step actions
在当前用户的 Exchange 草稿箱中创建 HTML 富文本邮件,但不发送。返回 draft_ref;后续优先使用 draft_ref 修改草稿或添加附件。
Production tool profile implementation for editing mail draft
使用结构化语义条件定位邮件,并隐藏 ItemId/ChangeKey。主 Agent 负责把"上周""昨天"等自然语言转换为 after/before。唯一结果返回 selected_message;多个结果返回 message_ref 候选。
7 of 25 tools (search_mail, read_mail, save_mail_draft, edit_mail_draft, continue_action, read_calendar, find_meeting_times, save_meeting, send_meeting_invitation) have only placeholder descriptions 'Production tool profile implementation for...' with no actual specification. These violate the baseline that 100% of A+ tools have substantive docstrings.
Input schemas are not explicitly declared as JSON Schema in the provided source code. Only Python function signatures with type hints are visible.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
Production tool profile implementation for finding meeting times
基于原邮件创建转发草稿,但不发送;优先使用 message_ref。
返回当前认证用户的显示名、主 SMTP 邮箱和 person_ref。如果无法从 NTLM 用户名唯一推断,会返回 ambiguous/not_found,并提示先用 CLI set-current-user 显式配置。不会暴露或返回密码。
读取一封邮件的正文、收发件人和附件元数据。item_id、message_ref、draft_ref 必须且只能提供一个。Workflow 应优先使用 ref。
内部兼容 helper;不注册为 Agent MCP tool。最多读取三周:最新一周返回为短 id 的 slots,前两周作为纯文本 history。HTML、内部 slot id、offset/hash、收件人等全部保留在 Server 的 resume_token 上下文中;本 helper 只读,不创建 Exchange 草稿。
按时间倒序列出邮箱中的邮件摘要并返回 message_ref/draft_ref。
Production tool profile implementation for reading calendar
Production tool profile implementation for reading mail
基于原邮件创建回复或全部回复草稿,但不发送。item_id/message_ref 二选一;优先使用 message_ref。ChangeKey 自动刷新。
按 message_ref 或搜索条件定位原邮件并创建回复草稿,绝不发送。
使用姓名拼音或完整邮箱查询个人联系人和全球通讯簿。不接受中文姓名查询。若 query 含非 ASCII 字符,将返回 romanized_query_required,由主 Agent 请求用户提供拼音或邮箱地址。
使用姓名拼音或完整邮箱解析人员,并结合历史邮件消歧。裸拼音会收集所有已配置公司域名下、邮箱 local-part 以该拼音开头的候选;不会优先选择无数字后缀的邮箱。多个候选中若只有一个存在历史来往才自动选择。
Production tool profile implementation for saving mail draft
Production tool profile implementation for saving meeting
搜索邮件摘要并返回稳定的 message_ref。folders 可跨多个标准文件夹搜索;设置 folders 时忽略 folder。participant_contains 会匹配发件人、To 或 CC;多个条件使用 AND。
Production tool profile implementation for semantic mail search
Production tool profile implementation for sending meeting invitation
修改现有草稿的收件人、主题、HTML 正文或重要性,绝不发送。item_id/draft_ref 二选一;优先使用 draft_ref。空列表可清空对应收件人字段。
处理"周报"请求的唯一 Agent 入口。request 是"本次要写入新周报的最终修改需求"。它既可以直接来自用户,也可以是 Agent 在调用本工具之前通过 search_mail/read_mail 读取其他人的周报或相关邮件后,先用 LLM 总结得到的自然语言输入。若用户要求"总结 A、B 发给我的周报并生成我的周报",应先完成邮件搜索、读取和总结,再把总结结果作为 request 调用 weekly_report;不要把"去搜索并总结 A、B"这类元指令原样作为 request。weekly_report 本身只处理自己的周报模板,不负责搜索他人的邮件。
Output schemas are not documented in the provided source code. Neither tools 1-16 nor tools 17-25 show explicit response structure documentation. Baseline: 100% of A+ tools have documented return types. This is a critical omission for LLM tool selection and chaining.
No error handling guidance is visible in the provided code. Tools do not document what errors they may return, how to recover, or whether errors are retryable. Baseline pattern (recovery-guide) requires: 'Error responses must tell the LLM what to do next.' This is missing across all tools.
Tool parameters lack consistent descriptions. Tools 17-25 have zero parameter documentation. Tools 1-16 have descriptions in mixed Chinese/English in the tool_profiles.py file, but the Python function signatures in server.py do not carry those descriptions inline. This violates the baseline: '94% of A+ tools have docstrings; 100% of A+ tool params have descriptions.'
Composition issue: tools like resolve_names, resolve_people, find_email, and reply_to_email perform similar name resolution and email lookup but with subtly different semantics. LLMs may conflate them. The server defines DEBUG_ONLY_TOOL_NAMES vs PRODUCTION_TOOL_NAMES, indicating internal tools are hidden, but the public surface still has overlapping semantic responsibility without clear differentiation in descriptions.
Some parameter descriptions lack format constraints. E.g., date parameters (after, before) in search_emails, find_email, reply_to_email are described as strings but no format (ISO 8601) is specified. The tool_profiles.py line mentions 'Agent 负责把"上周""昨天"等自然语言转换为 after/before' (Agent must convert natural language to after/before), but this responsibility distribution is not formalized in parameter descriptions.
Parameter naming inconsistency: Some tools use item_id, message_ref, draft_ref, change_key (e.g., get_email, reply_as_draft, update_draft) with mutual exclusion noted in descriptions, but JSON Schema constraints are not visible. Other tools (search_emails, find_email) do not expose these low-level IDs, relying instead on semantic queries. This inconsistency may confuse LLMs about which identifier type is expected.