Payper MCP has clear, action-oriented tool names (list_gpus, rent_gpu, check_session, stop_session) and solid descriptions (avg 180 chars, within baseline 194). All four tools have input schemas with Zod validation and parameter descriptions. However, output schemas are not formally documented, responses are shaped ad-hoc in the handler code (ok() helper returns untyped text). Error handling is present but generic (fail() wraps errors as text). No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk profiles. Parameter descriptions are good but lack explicit constraints (e.g., maxUsd has no min/max bounds stated). Overall solid for a small domain-specific server, but lacks production-grade output documentation and error guidance.

• Add formal output schema documentation for each tool. E.g., list_gpus returns {gpus: [{gpu: string, label: string, usdHour: number, available: number}], total: number}. Use JSDoc or TypeScript interfaces visible in code.
• Add tool annotations: rent_gpu and stop_session should have destructiveHint=true; list_gpu and check_session should have readOnlyHint=true. Use server.tool(..., {annotations: {readOnlyHint: true}}).
• Enhance error messages with recovery guidance. E.g., 'PAYPER_WALLET_KEY not set, set it to a 0x-prefixed private key of a wallet holding USDG on Robinhood Chain (chain 4663).' or 'GPU slug not found, call list_gpus() to see available options.'
• Document parameter constraints in descriptions. E.g., maxUsd: 'Maximum USDG to authorize (0.01 - 10000, typically 1 - 100 for short sessions)'. gpu: 'GPU slug from list_gpus output, e.g. rtx-3060, a100-sxm4, h100-pcie.'
• Add limit and pagination to list_gpus if catalog grows. Return {gpus: [...], total: number, limit: number} and accept optional limit parameter (default 50, max 200).