CryptoGuard exposes a single tool 'cryptoguard_check' with a well-defined input schema and reasonable description. However, the tool lacks output schema documentation, error handling guidance, and tool annotations. The naming is verb_noun compliant but the description, while detailed, could better explain WHEN to use this tool vs alternative security audits. Parameter descriptions are adequate but lack explicit format/constraint guidance. No idempotent, readOnly, or destructive hints are present. The implementation shows basic competence but falls short of production-grade agent tool standards (missing ~30% of expected quality markers).

• Add explicit outputSchema to the MCP tool definition. Document all response fields: risk_score (integer 0-100), risk_level (string enum), token_info (object with name, symbol, decimals), should_block (boolean), critical_findings (array of {title, description, severity}), high_findings (array), and summary (string). This enables LLMs to extract and chain results.
• Declare tool annotations in the MCP definition: set readOnlyHint=true and idempotentHint=true to signal safety. This is a 2026-07-28 spec alignment requirement that agents use for planning.
• Convert 'chain' parameter to an enum: instead of a default string, define enum: ['ethereum', 'bsc', 'polygon', 'arbitrum', 'base', 'optimism', 'avalanche', 'fantom', 'zksync', 'linea', 'scroll']. This prevents hallucinated chain names and makes the tool self-documenting.
• Add validation and error handling guidance to the description. Example: 'Returns error if address is invalid (must be 42-char 0x-prefixed hex), if chain is unsupported, or if all security APIs are unreachable. Partial results may be returned if some services time out. Queries may take 5-30 seconds for full analysis.' This prepares the LLM for real outcomes.
• Expand the 'address' parameter description: 'ERC-20 or token contract address on the specified chain. Must be a valid 42-character hexadecimal string starting with 0x. Example format: 0x1f9840a85d5aF5bf1D1762F925BDADdC4201F984. LLM agents: do NOT omit the 0x prefix; validate format before calling.'
• Document the response risk_level enum values in the description (e.g., 'CRITICAL, HIGH, MEDIUM, LOW, SAFE') so agents can condition on them reliably.