Loading ToolBench
An MCP server that provides access to external AI models through OpenRouter for debugging assistance, code optimization advice, and alternative coding approaches
Single tool with reasonable naming and description, but significant schema and composition gaps. The tool 'call_external_model' uses a verb start (call_) which is good, but the inputSchema is present and valid JSON Schema. However, the tool attempts to bridge multiple concerns (model selection, message passing, context inclusion) without clear parameter constraints. Description is 191 chars (within baseline 194 avg) and explains the purpose adequately. Parameters lack enums where they should exist (model: 'nano', 'mini', 'full' are hardcoded in description but not schema). Output schema is not documented in the tool definition, LLM cannot predict response structure. Error handling returns generic text wrapping rather than structured guidance. The tool is self-contained but does not compose well with other tools since no other tools exist.
When you hit a debugging roadblock or need a second opinion on code issues, consult external AI models. Use this for complex error debugging, optimization advice, or exploring alternative approaches to coding problems.
Model parameter accepts free-form string but description implies only 3 valid values ('nano', 'mini', 'full'). Should be declared as enum in inputSchema to prevent hallucinated model names.
Output schema not documented. Tool returns JSON with 'response', 'modelUsed', 'usage' fields, but LLM has no formal schema to understand the structure. Must add outputSchema or describe return type in tool description.
Error responses are plain text ('Error: <message>') wrapped in JSON. Should return structured error with actionable guidance, e.g., 'Model "xyz" not supported. Available: nano, mini, full. Try again with a supported model.'
No pagination or result limit documented. If external model returns large responses, context window could be exhausted. Should document max_tokens and offer truncation/pagination guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 57 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 42 | - | v1 |
Parameter 'message' is required but lacks format/length constraints in schema or description. Should specify: min length (e.g., 10), max length (e.g., 5000), and what constitutes a valid debugging question.
API token and endpoint are loaded from config file but not validated. If config is malformed or credentials are wrong, error handling is generic. Should validate config at startup with clear failure messages.