Single tool with minimal documentation. Tool name is action-verb compliant (perplexity_search), but description is extremely sparse (only 34 chars, well below the 194-char baseline average). Input parameter 'query' has a basic description but lacks guidance on format, constraints, or examples. No output schema documentation exists, the response structure is inferred only from the implementation (text field in content array). No error handling guidance for the LLM, raw HTTP and API errors are wrapped in generic McpError messages without recovery hints or actionable next steps. Security is acceptable (API key injected via environment variable), but the tool lacks composition guidance for multi-step workflows.

• Expand tool description to 100-150 chars. Example: 'Search the web using Perplexity AI. Use this to find current information, news, and web content when the user asks factual questions. Returns a single text response with search results and citations. Requires PERPLEXITY_API_KEY environment variable.'
• Document the output schema explicitly in code or README: 'Returns {content: [{type: "text", text: string}]}. The text field contains the full search response with inline citations.'
• Enhance query parameter description: 'The search query (max 1000 chars). Can include multiple terms, phrases in quotes, or boolean operators (AND, OR, NOT). Examples: "latest AI research", "python asyncio tutorial", "(machine learning OR deep learning) AND tutorial"'
• Add error handling with recovery guidance: Distinguish 401 (missing/invalid PERPLEXITY_API_KEY, user must set env var), 429 (rate limited, retry after backoff), 5xx (Perplexity service down, retry in 60s), and network timeouts. Return specific, actionable error messages.
• Document API constraints in description: 'Subject to Perplexity API rate limits. Queries timeout after 30 seconds. Best effort, may return partial results if API is unavailable.'
• Consider adding a 'max_results' or 'search_depth' parameter (optional, default='balanced') to let agents control latency vs result richness.
• Add tool annotations if supported: mark as readOnlyHint=true (does not modify state), and idempotentHint=true (same query always returns consistent results).