Three well-named tools (audit_html, audit_file, audit_url) with clear verb-noun structure and solid descriptions (150-200 chars each). All tools have input schemas with typed parameters and descriptions. However, output schemas are not documented, responses are returned as JSON text blobs without a formal schema definition. Parameter descriptions are good but lack constraint details (e.g., no guidance on valid tag values). Error handling is minimal, no recovery guidance or categorization. Missing tool annotations (readOnlyHint, idempotentHint). Composition is clean: each tool does one thing, but no pagination or result limits are enforced.

• Document the output schema formally: define AuditResult as a structured type with fields (violations[], passes, incomplete, inapplicable) and include it in tool metadata or a separate schema reference.
• Add tool annotations: set readOnlyHint=true and idempotentHint=true on all three tools to signal to LLMs that these are safe, repeatable operations.
• Constrain the 'tags' parameter: replace the generic description with an enum or pattern. E.g., 'One or more of: wcag2a, wcag2aa, wcag21a, wcag21aa, wcag22aa, best-practice, experimental, section508.'
• Implement error handling for audit_url: catch network errors, timeouts, and HTTP errors (4xx, 5xx). Return structured errors with recovery hints: 'Failed to fetch URL (timeout). Verify the URL is accessible and try again.'
• Add a result limit and document it: cap violations at 50 per response and note in the description: 'Returns up to 50 violations. For full audit results, use audit_html or audit_file with a real browser (Playwright/Puppeteer).' This prevents context bloat.