Loading ToolBench
A multi-tool MCP server that integrates CrewAI agents with web search, image generation, and filesystem access capabilities for comprehensive research and content creation
Three tools with basic registration, but critical gaps in definition quality. Tool names lack clear action verbs (image_creation_openai is confusing; brave_search and search_news are acceptable). Descriptions are minimal and under 200 chars, but descriptive text exists for all tools. Parameter schemas are present with types and descriptions, but lack constraints (enums, ranges, patterns). Output schemas are not explicitly documented, responses are inferred from code (success/error dictionaries). Error handling exists but is not actionable for LLM recovery. No tool annotations (readOnlyHint, destructiveHint). Overall definitions show basic competence but miss production-grade polish expected for agent coordination.
Search the web using Brave Search API
Create an image using Segmind API
Search for news using Brave Search API
Tool name 'image_creation_openai' is misleading, the tool uses Segmind API, not OpenAI. This violates the verb_noun_resource naming pattern and will confuse LLMs about which service is actually called. Should be 'create_image_segmind' or 'generate_image'.
No output schemas documented. Tool responses are dicts with 'success', 'error', 'image_path'/'results' fields, but these are only visible by reading implementation code. LLMs cannot plan downstream calls without knowing return structure. No way for agents to know if they should extract 'image_path' or another field.
Parameter constraints missing. The 'count' parameter in brave_search and search_news documents max=20 in description but has no JSON Schema minimum/maximum fields. LLMs will ignore undeclared bounds and may pass invalid values. Similarly, 'query' has no length limits or pattern validation.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 46 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 33 | - | v1 |
Error responses lack recovery guidance. brave_search returns {'success': False, 'error': 'BRAVE_API_KEY not configured'}, but does not tell the LLM what to do next. Should suggest: 'Set BRAVE_API_KEY environment variable and retry' or link to prerequisite setup.
No tool annotations. image_creation_openai is destructive (writes files, calls external API) and irreversible, but has no destructiveHint annotation. brave_search and search_news are read-only and idempotent, but lack readOnlyHint and idempotentHint. Agents cannot reason about safety without these hints.
search_news returns only 'results' field in top-level dict; brave_search also returns 'query' and 'total_results'. Inconsistent response schemas across similar tools force LLMs to handle different shapes and waste reasoning cycles. Canonicalize response structure.