Fully local MCP server and CLI for web research with search-first architecture using SearXNG for discovery, Crawl4AI for extraction, and MarkItDown for document conversion
Three well-named, verb-led tools (search_web, read_pages, read_urls) with comprehensive descriptions (150-250 chars each). All parameters have types and descriptions. Input schemas are complete with enums for 'effort' parameter. However, output schemas are not documented in the visible code, only input schemas are defined. Error handling guidance is minimal. Tool composition is sound: search_web discovers and caches, read_pages/read_urls retrieve cached content. Descriptions are LLM-optimized with clear use-case guidance (quick/normal/deep effort levels). No security issues detected (no secrets in params). Naming follows verb_noun convention consistently.
Retrieve cleaned content for one or more stored pages using page_ids returned by search_web. Batch related page_ids in one call when comparing multiple sources.
Retrieve cleaned content for one or more direct URLs without running search_web first. Supported document URLs such as PDFs are converted automatically when detected.
Search the web for relevant sources and return compact summaries with stable page_ids for follow-up reads. Choose effort deliberately: quick for narrow current-fact lookups or explicit URLs, normal for most docs and troubleshooting, and deep for broad comparisons or synthesis-heavy research. Use domains when you want to constrain results to specific hosts.
Output schemas not documented. Tool descriptions explain what is returned (summaries, page_ids, cleaned content) but formal output schema definitions are absent from visible code. LLMs cannot plan downstream operations without knowing return field names and types.
Error handling lacks recovery guidance. No documented error responses or categorization (retryable vs fatal). Agents cannot determine whether to retry, ask user, or abort on failures like timeouts, 404s, or blocked pages.
Pagination not addressed. search_web may return many results but no limit, offset, or next_cursor parameters are visible. Large result sets risk context window exhaustion.
Tool annotations missing. No readOnlyHint, destructiveHint, or idempotentHint declarations visible. All three tools are read-only (safe to retry), but this is not formally signaled to clients.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 70 | 2026-07-28+ | v2 |
Parameter 'focus' lacks format guidance. Described as 'short topic phrases, exact errors, function names' but no length limits, regex patterns, or examples provided. LLMs may pass overly long or malformed focus strings.