Project Door demonstrates a sophisticated access-control architecture but falls short on tool definition quality. All 4 tools have descriptions and basic schemas, but parameter descriptions are entirely absent. The ping tool is a deprecated MCP pattern (removed in 2026-07-28). Input schemas lack parameter-level type information and constraints. Descriptions are terse (10-100 chars) and do not explain WHEN to use each tool or what downstream actions they enable. No output schemas are documented. Error handling is implicit in the guard logic but not surfaced to the LLM. The server prioritizes security policy over agent usability.

• Add parameter descriptions to all schemas. For 'amount', specify: 'Payment amount in USD cents (integer, 1 - 1000000). Requires live human presence to authorize.' For 'to', specify: 'Wire recipient account number (10 - 20 digits) or email address. Must match treasury whitelist.'
• Document output schemas. E.g., list_invoices should return {invoices: [{id, amount, date, status}], total_count}. send_payment should return {transaction_id, timestamp, authorized_by_principal}. This enables chaining and structured parsing.
• Remove the ping tool. It is deprecated in MCP 2026-07-28. If health checks are needed, use a GET /.well-known/health endpoint outside the tool interface.
• Add tool annotations (destructiveHint, idempotentHint) to signal to clients which tools are safe to retry. Mark send_payment and wire_transfer with destructiveHint=true so agents know these are irreversible.
• Embed recovery guidance in error responses. Instead of returning HTTP 401 with www-authenticate, return a structured error: {error: 'biometric_required', message: 'Wire transfer requires device biometric. Unlock your device and retry.', retryable: true}.