Loading ToolBench
mcp-auditor is a security testing tool, not a production MCP server. The 8 honeypot tools are intentionally flawed for evaluation purposes (planted SQL injection, info leakage, missing validation). Naming is verb-based (project_manager, user_directory, get_user, execute_query, list_items, search_users, delete_record, get_service_status), but descriptions are minimal (10-50 chars) and lack LLM-optimized guidance. Schemas are present with enums and type constraints, but parameter descriptions are sparse. Error handling is deliberately weak per design. This is a test harness, not a production tool suite.
Delete a record by ID.
Execute a SQL query against the database.
Check the health status of a service.
Look up a user by their numeric ID.
List items in a given category with an optional limit.
Manage projects: list, view details, or read project files.
Search for users by name.
Descriptions are under 50 characters for 6/8 tools, violating the 10 - 1024 character baseline. LLMs cannot determine when to select tools without context.
Parameter descriptions are minimal or missing context. E.g., 'action' enum values are listed but not explained (what does 'read_file' do vs 'details'?). LLMs cannot infer intent from names alone.
No output schema documentation. Tools return JSON strings, but LLMs don't know the structure (fields, types, nesting). Agents must guess or parse unstructured text.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
Search users, view profiles, or retrieve specific fields.
Error handling is deliberately weak (planted flaws: raw KeyError, info leakage in error messages, no recovery guidance). Production servers must guide LLMs on retry vs user-fixable vs fatal errors.
delete_record lacks confirmation or dry-run support. A destructive operation should require explicit confirmation to prevent accidental data loss by agents.