Loading ToolBench
Demo of how an ExpressJS server can be used as a LLM MCP server.
Scoring was not performed
Parameter description for targetNumber lacks explicit constraints. Should state 'integer, range 1 - 1025, defaults to 25 (Pikachu)' but instead embeds validation rules in prose prose, forcing LLMs to parse unstructured text.
No output schema is documented for either tool. LLMs cannot plan downstream calls or extract specific fields. The description says 'data about the Pokemon' but does not structure the expected response fields (e.g., {name, id, types, height, weight}).
get_pokemon_by_number description is 303 characters, exceeds optimal 200-char baseline and buries the 'defaults to Pikachu' detail in a complex conditional. Should be split: 'Retrieves detailed Pokémon data by National Pokédex ID (1 - 1025). Defaults to Pikachu (ID 25) if no ID provided.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 20 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 36 | - | v1 |
No error handling guidance in tool definitions. If targetNumber is invalid (e.g., 9999 or -1), the tool returns a plain error string. Definitions should state: 'Returns error if targetNumber is not between 1 and 1025, or if the PokeAPI request fails. Suggest retrying with a valid ID.'
Tool descriptions do not explicitly state these are read-only operations (no side effects, safe to retry). Agents need this classification to plan safely and avoid unnecessary idempotency guards.
targetNumber parameter is a number type but lacks min/max JSON Schema constraints. Should include 'minimum': 1, 'maximum': 1025 in the schema to enforce bounds at parse time rather than runtime.