This server has critical structural and definition gaps. Tool definitions exist only as string names in a manifest; no actual MCP tool registration with schemas is visible in the source code. The three tools (issue_agent_card, execute_purchase, verify_spend_limits) are declared but not implemented as callable MCP tools. Input schemas are inferred from the problem statement but not present in the actual codebase. The mcp_server.py file only prints a static manifest and does not handle tool calls, parameter validation, or error responses. Client.py contains mock implementations that return hardcoded responses, not real integrations. No tool descriptions, parameter constraints, or output schemas are enforced in the server logic.

• Implement proper MCP tool registration in mcp_server.py using the MCP SDK. Each tool must be registered with a name, description, input schema (JSON Schema with types and descriptions for all parameters), and a handler function.
• Add JSON Schema definitions for all tool inputs. For issue_agent_card: amount_usd (type: number, min: 0.01, max: 10000), merchant (type: string, pattern for domain), purpose (type: string, min 10 chars), card_type (enum: ['single_use', 'multi_use']). For execute_purchase: ask (type: string, min 10 chars), max_budget_usd (type: number, min: 0.01).
• Implement verify_spend_limits with a schema (card_id: string) and description: 'Retrieve current spend limits and remaining balance for an issued virtual card.'
• Move API credentials to server-side environment variables (AGENTCARD_API_KEY, AGENTCARD_BASE_URL). Remove api_key from AgentcardClient constructor and client.py. Inject secrets server-side only.
• Add error handling with recovery guidance. Return structured errors: {"error": "card_limit_exceeded", "message": "Spend limit $X exceeded. Request a new card with higher limit.", "retryable": false}. Classify errors as retryable (network timeout) vs user-fixable (invalid merchant domain) vs fatal (permission denied).