MCP server for SAP automation providing tools to create sales orders, stock transfer orders, delivery notes, materials, source lists, info records, and manage kitting flows with SAP system integration
This server has severe definition quality gaps across all dimensions. Tool names lack consistent action verbs (e.g., 'set_sap_credentials' is acceptable, but 'create_sales_order' lacks domain-specific context). Descriptions are present but often generic and under 100 characters, providing minimal guidance on WHEN or WHY to use each tool. Input schemas are defined with proper JSON Schema types and descriptions, which is a positive. However, no output schemas are documented, the LLM cannot understand what fields to expect from responses, breaking downstream tool composition. Error handling is absent: the code returns raw HTTP errors and XML parsing failures without actionable recovery guidance. The server stores credentials per-session in memory, but exposes them via tool parameters (set_sap_credentials takes plaintext user/password), violating the secret-injection pattern. No pagination, rate limiting, or permission gates are visible. Tool descriptions are in Traditional Chinese, which is fine for non-English deployments, but the descriptions are terse (avg ~80 chars) and lack dependency hints or prerequisite context.
在 SAP 系統中變更套件流程數量(Kitting Flow Quantity)
檢查目前工作階段是否已設定憑證 回傳: 憑證狀態資訊
在 SAP 系統中建立出庫單(Delivery Note)
在 SAP 系統中建立或維護供應商資訊記錄(Info Record)
在 SAP 系統中建立新物料
在 SAP 系統中建立銷售訂單
在 SAP 系統中建立或維護供應商源列表(Source List)
在 SAP 系統中建立庫存轉移訂單(Stock Transfer Order)
Credentials exposed as tool parameters instead of server-side injection. set_sap_credentials takes plaintext username and password as inputs, which violates secret-injection pattern and risks credentials leaking into agent traces, logs, and prompt history.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract chaining IDs (e.g., if create_sales_order returns a sales_order_id, create_delivery_note needs that ID). This breaks tool composition.
Tool descriptions are too brief (avg ~31 chars) and lack context on prerequisites, return structure, and when to use each tool. Below the 50-200 char LLM-optimized baseline (rubric baseline: 194 chars p50).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 38 | 2026-07-28+ | v2 |
查詢 SAP 系統中訂單流程的狀態
設定目前工作階段的 SAP 憑證 參數: username: SAP 使用者名稱 password: SAP 密碼 回傳: 設定結果訊息
No error handling or recovery guidance. Code returns raw HTTP status codes and XML parsing exceptions (e.g., 'HTTP Error 200: ...' or 'Connection Error: ...'). No actionable messages to guide LLM on retry vs. user-fixable issues.
Several input parameters lack validation constraints or enums. E.g., 'material_type' (no valid values), 'status' (enum hidden in description, not in schema), 'new_quantity' (no > 0 constraint). Parameters named identically in input but no validation that they exist in SAP (e.g., does customer_id in create_sales_order exist?).
No pagination support. Tools like query_flow_status (and any list tool, if they existed) lack page/limit/offset parameters. No indication of result size limits. Large result sets risk context window exhaustion.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLM cannot infer whether a tool modifies state (e.g., create_* tools) without these hints. This violates current MCP spec best practices.
Session credential management is memory-resident only. No persistence. If the server restarts, all credentials are lost. No clear documentation of session scope or lifetime.