Loading ToolBench
Two tools with significant quality gaps. 'research' has a reasonable description (89 chars) and structured input schema with enums, but lacks output schema documentation and error handling guidance. 'ping' is a deprecated anti-pattern (removed from MCP spec 2026-07-28) with minimal description (54 chars). Neither tool documents what fields are returned, how to chain results, or how to recover from failures. Parameter descriptions are present but generic. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Missing pagination guidance for 'research' results.
Health check — returns ok if the main agent can answer.
Run a multi-source research task. Decomposes the prompt, dispatches parallel sub-agents across the requested sources, then synthesizes an answer.
ping tool is a deprecated anti-pattern removed from MCP spec 2026-07-28. Health checks should use HTTP /health endpoint, not a tool.
No output schema documented for 'research' tool. LLMs cannot plan downstream calls or extract required fields (e.g., source citations, confidence scores, next_cursor for pagination).
Missing error handling guidance. No description of what errors can occur, whether they are retryable, or what the LLM should do next (e.g., 'If depth=deep times out, retry with depth=standard').
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). 'research' is read-only but this is not declared in the schema, forcing LLMs to infer safety.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | <=2025-11-25 | v2 |
Parameter 'scope' accepts an array of strings with no enum constraint or description of valid values. LLMs will hallucinate source names instead of selecting from a known set.