Loading ToolBench
MCP server for Perplexity AI API integration
Two tools with clear action verbs (perplexity_ask, perplexity_search) and good descriptions (180-220 chars, within baseline 194 avg). Input schemas present with proper JSON Schema structure, enums for model selection, and defaults. However, parameter descriptions are minimal (10-30 chars vs 72 char baseline), lacking guidance on when to use each model or how max_tokens affects output. No output schema documented, LLMs cannot predict response structure. Error handling returns text but lacks recovery guidance or categorization. No tool annotations (readOnlyHint, idempotentHint). Rate limiting is server-side but not exposed to LLM.
Send a direct question to Perplexity AI and receive a comprehensive answer. This tool leverages powerful AI models to analyse and respond to complex questions with detailed, factual answers. Citations for sources are included when available. Best for direct questions requiring factual or analytical responses.
Perform a web search query with Perplexity AI to find relevant information online. This tool combines web search capabilities with AI-powered analysis to deliver comprehensive search results with source citations. Best for research questions, current events queries, or when you need information from multiple online sources.
Parameter descriptions are too brief (10-30 chars vs 72 char baseline). 'Perplexity model to use' and 'Maximum number of tokens in the response' lack guidance on when to choose sonar-reasoning vs sonar-pro, or how token limits affect quality/cost tradeoffs.
No output schema documented. Tool responses return {content: [{type: 'text', text: '...'}], isError: boolean} but LLMs cannot predict this structure. Agents cannot extract citations, token usage, or model metadata for downstream reasoning.
Error responses lack recovery guidance. When rate limit is exceeded or API fails, the tool returns 'Error: <message>' as plain text. LLMs cannot determine if the error is retryable, user-fixable, or fatal.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
No tool annotations. Both tools are read-only (marked in metadata) but lack readOnlyHint in schema. Agents cannot infer safety properties without explicit annotations.
system_prompt parameter is exposed and mutable. While not a secret, allowing LLMs to override system instructions risks prompt injection. Consider making this server-side only or documenting strict constraints.