serp-it has three well-defined tools with clear names (search, search_fetch, fetch) and comprehensive descriptions (130-180 chars each). All tools have documented input schemas with proper type constraints (minLength, maxLength, optional flags). However, output schemas are inferred from Zod definitions in code rather than explicitly exposed in tool registration, creating ambiguity about what LLMs should expect. Tool annotations (readOnlyHint) are present on all three tools, which is excellent. Parameter descriptions are thorough and include usage guidance (e.g., 'Use quotes for exact phrases'). The resource capability (serp-it://capabilities) and three prompts (research, fact-check, compare-sources) add value but don't substitute for tool-level clarity. Main gaps: (1) Output schemas not directly visible in tool registration payloads, LLMs must infer structure from execution; (2) No explicit error handling guidance in tool descriptions (what happens on network failure, invalid URL, unsupported MIME type?); (3) search_fetch's maxFetch parameter lacks guidance on timeout/performance implications of large fetches; (4) No idempotent/confirmation patterns for potentially expensive operations.

• Expose output schemas explicitly in tool registration: add a 'result_schema' or document the response structure in MCP tool metadata. LLMs need to see that search() returns {results: [{title, url, snippet, sources}], engineErrors?: []}, not infer it.
• Add error handling guidance to tool descriptions. E.g., search: 'Returns empty results if all engines fail; inspect engineErrors array for details.' fetch: 'Returns empty markdown and error in response if rendering fails; supports HTML and PDF; times out after 30s.'
• Replace region parameter's loose string type with enum of supported locales. Define as: enum=['en-US', 'en-GB', 'fr-FR', 'de-DE', 'ja-JP', 'es-ES', ...] based on what the underlying engines support.
• Document search_fetch timeout and practical fetch limits in description: 'Fetches top N results; each fetch has 30s timeout; recommend maxFetch ≤ 10 for reliability.'
• Add parameter for fetch() to control markdown rendering options (e.g., include_images: boolean) and document max response size (e.g., 'Markdown output capped at 40KB; truncated if larger').
• Consider adding a 'dry_run' or validation mode for fetch() to check URL accessibility before expensive rendering.
• Document what constitutes failure for each tool and how to detect partial success (e.g., search returned 3 results from 7 engines, is that success or partial failure?).