A Model Context Protocol server that provides web search and research capabilities using Perplexity AI's API
The server defines a single tool 'search_web' with a clear verb-noun name and good semantic structure. The tool description is clear and appropriately detailed (114 chars). The input schema is properly structured with type definitions and descriptions for all parameters. However, there are significant gaps: (1) no output schema documentation, the tool returns text content but doesn't describe the structure or fields agents should expect; (2) parameter descriptions, while present, lack actionable format guidance (e.g., what constitutes a good query, length limits, whether free-form or structured); (3) no error handling guidance or recovery patterns in the tool description; (4) the optional parameters (type, complexity) have good enum constraints but lack hints about when to use each variant; (5) no documentation of response format, making it harder for LLMs to chain results into subsequent tools. The implementation shows sophistication (escalation logic, role-based system prompts), but these details are not surfaced in the tool definition where agents need them.
Research technical topics, troubleshoot issues, or get latest updates using Perplexity AI
No output schema documented. Tool returns text content via OpenAI API but the definition provides no guidance on structure, formatting, or what fields/metadata LLMs should expect in responses.
Parameter descriptions lack actionable format constraints. 'query' param has no guidance on minimum length, expected complexity, or examples of good vs. poor queries. 'type' and 'complexity' enums are present but descriptions don't explain when to use research vs troubleshoot vs update.
Tool description omits error handling and recovery guidance. Does not state what happens on API failures, rate limits, or invalid queries. Agents have no instruction on when to retry vs. escalate vs. ask user.
No indication that tool modifies state or has side effects. While search_web is read-only, the description should explicitly state this so agents know the call is safe to retry without consequence.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 7 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 47 | - | v1 |