UnSearch MCP Server has 4 tools with reasonable naming and solid parameter schemas, but falls short in several critical areas. Tool descriptions are present and moderately detailed (90-180 chars), meeting baseline length requirements. All tools have input schemas with typed parameters and descriptions. However, output schemas are completely undocumented, the tool descriptions mention returned data (citations, verdicts, confidence scores) but the actual response structure is not specified anywhere in the provided source. Error handling guidance is minimal. The server lacks tool annotations (readOnlyHint, destructiveHint, idempotentHint) required by current spec. Parameter descriptions are adequate but could be more actionable (e.g., 'engines' parameter defaults are listed but rationale for those defaults is not explained).

• Document the output schema for each tool. For 'search': return an array of objects with {url, title, snippet, signed_citation: {url, fetched_at, content_sha256, snapshot_key, signature}}. For 'extract': return {raw_text, markdown, citation}. For 'research': return {answer, sources: [{title, url, excerpt, citation}]}. For 'verify_claim': return {verdict, confidence: 0.0-1.0, evidence_spans: [{text, citation}]}.
• Add readOnlyHint: true to all four tools in the tool definitions. This signals to agents that these are safe, non-destructive calls.
• Constrain numeric parameters: 'max_results' (search) → min 1, max 20 (align with description); 'max_sources' (research) → min 1, max 50.
• Add error recovery guidance to tool descriptions. E.g., 'search': 'If a result URL fails to load, the agent should try the next ranked result or use a different search engine.'; 'extract': 'If extraction fails due to paywall or robots.txt, returns null content with explanation; agent should skip that URL.'
• Explain 'depth' parameter tradeoffs in descriptions. E.g., 'research': 'depth=quick (1-2 sources, <5s); depth=standard (5-10 sources, ~10s); depth=deep (15+ sources, ~30s); depth=comprehensive (all relevant sources, may exceed 60s).'
• For 'engines' parameter, list all available engines (google, bing, duckduckgo, etc.) and note that some engines may have regional restrictions or require authentication. Suggest default engines are fastest; custom engines for specialized queries.
• Add 'include_analysis' and 'include_summary' parameter descriptions explaining what analysis entails (trend detection, source credibility assessment) and when to disable them (to reduce latency for simple lookups).