Crawl4AI MCP has basic tool definitions but significant quality gaps. Both tools have descriptions and input schemas, but descriptions are generic (under 100 chars), parameter descriptions lack actionable detail, and output schemas are not documented. The crawl_website tool accepts many parameters but lacks constraints (e.g., depth range, maxPages bounds). process_text exposes model selection as a parameter, which is unusual for an MCP server. No error handling guidance, no idempotency hints, and no security considerations documented. The server is HTTP-based (good), but tool definitions read as API wrappers rather than LLM-optimized interfaces.
Crawls a website starting from the specified URL
Process crawled text content with AI
Output schemas not documented. crawl_website returns an array of page objects with url, title, description, content, html, crawledAt fields, but this structure is not declared in the tool definition. process_text output is completely undocumented. LLMs cannot plan downstream tool calls without knowing what fields to expect.
Parameter descriptions lack actionable constraints. 'depth' is described as 'How many levels deep to crawl (default: 1)' but no min/max bounds stated. 'maxPages' has no range. 'waitTime' in milliseconds has no guidance on reasonable values. LLMs will pass arbitrary numbers without constraints.
process_text exposes 'model' as a parameter (default: claude-3-sonnet-20240229). This is unusual for an MCP server, model selection should be server-side configuration, not a tool parameter. Exposing it invites LLMs to hallucinate model names or override intended defaults.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | <=2025-11-25 | v2 |
No error handling guidance. If crawl_website fails (invalid URL, network timeout, robots.txt block), the tool returns a generic error. No recovery hints like 'Try a different URL' or 'Check if the domain is accessible'. Agents cannot self-correct.
No idempotency or destructiveness hints. crawl_website is read-only (good), but process_text modifies external state (calls Claude API, incurs costs). No tool annotations (readOnlyHint, destructiveHint, idempotentHint) to guide agent retry logic or cost awareness.