Simulates payment workflows, fraud scoring, and consensus voting with multi-agent integrity checks and transaction evaluation.
The MCP Payments Simulator demonstrates poor definition quality across multiple dimensions. While all 5 tools are explicitly registered with the @mcp.tool() decorator and have descriptions, critical gaps exist in schema completeness, parameter descriptions, and error handling guidance. Tool names follow action-verb conventions (create_, get_, execute_with_, score_), but descriptions lack depth and recovery guidance. Parameter schemas are present but inconsistently detailed. No tool output schemas are documented, forcing LLMs to infer response structures. Error handling is minimal, most tools return generic error strings without actionable guidance or recovery paths. The consensus-voting tool (execute_with_consensus) has concerning descriptions about 'revocation' and 'tampering' but no security-focused error messages or permission gates. Overall, this reads as a simulation/demo server rather than a production-grade tool set.
Creates a merchant-locked virtual card with a spending limit.
Executes a transaction via multi-agent consensus voting with integrity checks. This tool evaluates agent integrity before consensus. Agents failing integrity checks are automatically revoked and excluded from the voting pool.
Provides a concise fraud risk score for a transaction.
Retrieves simulated recent receipts for a customer.
Scores a payment for fraud risk and provides a recommendation.
Duplicate tools with identical functionality: get_fraud_score and score_payment_risk both accept (amount, merchant, hour) and return fraud scores. No description distinguishes them or explains when to use each. Violates single-responsibility and creates LLM confusion.
No output schemas documented for any tool. Code returns formatted strings (create_merchant_locked_card), nested report structures (execute_with_consensus), and dicts (fraud scoring), but LLMs cannot know structure ahead of time. Violates pattern:tool and forces inference-based reasoning.
Missing parameter constraints: 'amount' parameters lack min/max bounds; 'hour' parameter lacks explicit 0 - 23 constraint in schema (only in code comments). 'days' in get_receipts has no documented range.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 46 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Error messages do not guide recovery. Examples: 'Card creation BLOCKED: Fraud score too high' (no suggestion to try lower amount or different merchant), 'Transaction BLOCKED: All consensus agents are currently revoked' (no actionable next step), 'Error: Amount must be positive' (bare error, not structured).
Security-critical tool (execute_with_consensus) lacks permission gates, scope declarations, and audit-trail documentation. Tool claims to 'revoke agents' and check 'integrity' but has no declared permissions (e.g., 'write:consensus', 'admin:revoke'). No audit logging declared. State management (_tampered_agents, db.revoke_agent) is undocumented.
Insufficient descriptions for domain-specific terms. 'Consensus voting', 'integrity checks', 'dual-signal detection', 'behavioral drift', 'revocation', 'tampering' are used in execute_with_consensus without explaining what they mean, when they trigger, or how LLMs should interpret them.
Tool descriptions are 10 - 70 characters for most tools (get_fraud_score: 57 chars, score_payment_risk: 58 chars), below the 100-char minimum for actionable LLM guidance. Lack context on prerequisites, side effects, and selection criteria.
No pagination or limit enforcement in get_receipts. Code simulates 3 random receipts, but no limit parameter, no total count, no next_cursor. If this were backed by real data, an LLM could request receipts for 365 days and receive thousands of items, exhausting context.