Loading ToolBench
A multi-agent orchestration system with MCP servers for Gmail, Google Calendar, Knowledge management, and personal notes. Includes Slack bot integration and dashboard.
Argus has 18 tools with mixed quality. Naming is generally verb-first and clear (compose_email, send_email, create_event, etc.), following conventions well. However, descriptions vary significantly: some are detailed (e.g., create_event with ISO8601 format guidance), while others are minimal or in Japanese without English translation (compose_email, send_email, search_lessons). Input schemas are present for all tools with proper JSON Schema structure, but parameter descriptions are inconsistent, some parameters lack detail on constraints, formats, or valid ranges. Output schemas are not documented anywhere in the visible code. Error handling is absent from tool definitions. The server lacks tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite having tools with clear risk profiles (WRITE, DESTRUCTIVE). No pagination support visible for list tools. Overall, the server is functional but falls short of production-grade quality due to incomplete descriptions, missing output documentation, and lack of error guidance.
新規メールのドラフトを作成します。送信はユーザーがSlackのボタンで確認後に行われます。メールアドレスが不明な場合はユーザーに確認してください。
Create a new Google Calendar event. Use ISO8601 format for dateTime (e.g. 2026-03-15T19:00:00+09:00) or YYYY-MM-DD for all-day events.
Delete a Google Calendar event by its ID.
Add a new knowledge entry. Requires collector role.
Archive (delete) a knowledge entry. Requires collector role.
List all knowledge entries ordered by last updated date.
Search knowledge entries by name or content. Returns matching entries.
Descriptions in Japanese without English translation (compose_email, send_email, search_lessons) violate the pattern:tool-description requirement. LLMs trained primarily on English cannot reliably parse Japanese descriptions, leading to incorrect tool selection.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields without knowing what each tool returns. This violates pattern:tool and pattern:response-shaper.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk profiles. delete_event, knowledge_archive, send_email are marked DESTRUCTIVE/WRITE but lack annotations to guide agent behavior.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Update an existing knowledge entry. Requires collector role.
List Google Calendar events in a time range. Returns up to maxResults events ordered by start time.
Create a new personal note file.
Get personal context for AI personalization. Sections: identity, values, strengths, thinking, preferences, routines.
List all personal notes, optionally filtered by category (e.g. 'self').
Read a specific personal note by path (relative to data directory, e.g. 'self/values.md').
Search personal notes by keyword. Returns matching lines with context from all markdown files.
Update an existing personal note. Use 'append' to add content or 'replace' to overwrite.
過去の実行から学んだ教訓を検索します。失敗が繰り返される場合に関連する教訓を探すために使用してください。
メールを即座に送信します。確認ボタンなしで直接送信されるため、インボックスなど明示的な指示がある場合に使用してください。
Update an existing Google Calendar event. Only specified fields are changed.
list_events and knowledge_list lack pagination parameters (limit, offset, page_size). Returning unbounded results risks context window exhaustion and violates pattern:paginated-result.
No error handling guidance in tool definitions. Tools lack recovery hints (e.g., 'If event not found, try list_events first'). This violates pattern:recovery-guide and forces agents to guess next steps on failure.