Loading ToolBench
LIME reference MCP door to PostgreSQL — agent passport identity first
Solid foundation with clear verb-noun naming (list_*, get_*, select_*, execute_*) and consistent schemas across 8 tools. All tools have descriptions (avg ~60 chars, baseline 194) and proper JSON Schema with types. Input parameters are well-typed with required fields marked. However, descriptions are terse and lack LLM-optimized guidance on WHEN to use each tool, dependencies, and recovery paths. Parameter descriptions are minimal or absent. Output schemas are not documented. No tool annotations (readOnlyHint/destructiveHint). Error handling guidance is absent from definitions.
Execute a single READ-class SQL statement.
Execute a single DML write statement.
EXPLAIN a single SQL statement.
Database size / admin stats.
Describe columns of a table.
List non-system PostgreSQL schemas.
List tables in a schema.
Descriptions are terse (avg ~50 chars vs baseline 194) and lack LLM-optimized guidance. Missing WHEN to use each tool, dependencies between tools (e.g., 'call list_schemas first'), and recovery hints. E.g., 'List non-system PostgreSQL schemas' does not explain when an agent should call this vs list_tables.
Parameter descriptions are missing or minimal. E.g., 'schema' and 'table' parameters in get_table_schema have no descriptions explaining what they accept, format constraints, or examples. LLMs cannot infer parameter meaning from names alone.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Select rows from a table (parameterized).
Output schemas are not documented in tool definitions. LLMs cannot plan downstream calls or extract required fields (e.g., what does select_rows return? Column names? Data types? Row count?). This forces agents to guess or make extra discovery calls.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). The execute_write_query tool is destructive but unmarked. Agents cannot distinguish safe (read-only) from risky (write) operations without explicit hints.
Error handling guidance is absent from tool definitions. No recovery hints (e.g., 'If SQL parse fails, check syntax' or 'If permission denied, verify role'). Agents receive errors but cannot self-correct or know what to try next.