MCP Server with web search capability
Single tool with basic functionality but significant definition quality gaps. Tool name 'search_web' follows verb_noun convention correctly. Description is present but generic (27 chars, below the baseline 194 avg). Input schema exists but minimal. Output schema is undocumented, LLMs cannot predict return structure. Error handling is present but provides only generic messages without recovery guidance. No parameter type annotation visible in schema definition (uses z.string() which does provide type, score adjusted). The tool exposes API key management through sessionId injection rather than proper parameter validation. Overall aligns with a D-range server with basic setup but lacks polish for production use.
Search the web for the given query
Output schema not documented. Tool returns structured text but LLMs have no way to predict the response format (organic results array, snippet fields, etc.). This violates the pattern requirement that tool output schema be documented so LLMs can plan downstream actions.
Description is 27 characters ('Search the web for the given query'), below baseline 194 chars (p10=34, p90=392). Lacks context for LLM selection: When should this be called vs a knowledge base lookup? What quality are results? Required? Optional? Does it call Serper API or another backend?
No result limits or pagination declared. Tool description does not state that results are capped at 3 (hardcoded in code). If Serper API returns more results, LLMs cannot know whether truncation occurred.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 46 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 45 | - | v1 |
Generic error handling. Catch-all 'Error: Could not complete the search.' does not tell LLM what to do next: Is the API down? Was the query invalid? Should it retry, ask user, or try a different approach? Per pattern: 'Error responses must tell the LLM what to do next.'
API key management via session headers is unconventional. Tool requires X-Serper-Api-Key header on initial SSE connection, stored in sessionApiKeys map, then referenced by sessionId inside tool execution. This is non-standard for MCP and couples tool execution to session state. Per pattern:secret-injection, credentials should be injected server-side at initialization, not passed per-connection. Current design makes key management opaque to LLM and complicates debugging.
Parameter description missing. The 'query' parameter has description 'The search query to look up on the web' (24 chars), which is minimal but present. However, constraints are not documented: Is there a character limit? Can it include special operators? Does it support boolean logic?