A mock agent marketplace system for managing item sales, price negotiations, and budget tracking between agents
This is a minimal MCP server with severe definition quality gaps. The codebase shows 3 tools (browse, propose, buy) implemented as command-line argument handlers in agent_market.js, but the tool registration mechanism, schemas, and descriptions are either missing or inferred rather than explicitly visible in the provided source. Tool definitions appear to be CLI-driven rather than properly registered as MCP tools with formal schemas. Descriptions are present but minimal (10-80 chars). Input schemas are partially visible only in the eval brief, not in the actual source code provided. No output schemas are documented. Error handling is present but returns raw error messages without recovery guidance. The code reads/writes a JSON file directly, suggesting a single-process, file-based implementation, not a proper MCP server with remote transport capability.
Display current available items in the agent marketplace and show the buyer's budget
Execute a purchase transaction: deduct price from buyer budget, add to seller budget, and mark item as sold
Propose a price for an item and receive seller acceptance or rejection based on negotiation logic
No explicit MCP tool registration visible in source code. Tools are implemented as CLI argument handlers (process.argv[2]), not as formal MCP ToolDefinition objects with schemas.
No input schemas defined in source code. Input parameters (itemId, buyer, price) are documented only in the eval brief, not in schema definitions visible in agent_market.js. Per HARD SCORING RULE, if schemas are not visible in source, score must be 0.
No output schemas documented anywhere. Responses are unstructured console.log statements. LLM cannot know what fields to expect (e.g., does browse return an array of items? What fields does each item contain?).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 37 | 2026-07-28+ | v2 |
browse tool description is extremely brief ('Display current available items in the agent marketplace and show the buyer's budget' = 84 chars). Missing context on WHEN to call it, what structure it returns, and WHEN to use it instead of propose or buy.
propose tool documentation does not explain when/why to call propose vs buy. LLM has no guidance on negotiation workflow or whether propose is idempotent.
No parameter descriptions for 'itemId', 'buyer', 'price' visible in source code (only in eval brief). Parameters lack type constraints (enums, ranges). E.g., 'price' is an integer but no min/max specified.
Error handling returns raw error strings ('Item not found.', 'Insufficient budget.') without actionable recovery guidance. Per pattern:recovery-guide, errors must tell the LLM what to do next (e.g., 'Item not found. Try browse() to see available items.').
No idempotency guarantees. buy tool modifies state (budgets, item.status) without transaction semantics. If buy is called twice with the same parameters, second call will fail ('Insufficient budget' after first purchase). No idempotent-operation pattern implemented.
hardcoded agent name 'agent_main' in browse output. Tools do not accept a 'user' or 'agent' context parameter, so the server cannot serve multiple agents. Tool cannot be called with different buyer/seller identities dynamically.
No pagination support. If the marketplace grows to 100+ items, browse returns everything, potentially exhausting context window. Per pattern:paginated-result, tools returning lists should support limit and offset/cursor.