Google Gemini SDK-based testing script for GAP MCP server with structured data support, guardrails, and multilingual agricultural chatbot capabilities
This is a test harness, not a production MCP server. The two tools (get_gap_weather_forecast, get_accuweather_weather_forecast) are defined in utils/gemini-mcp-tool.js but lack critical production-grade structure. Descriptions are generic (identical for both tools, 54 chars). Parameter descriptions exist but are minimal. No output schema is documented. The codebase is a Gemini SDK testing script (batch-test-sdk.js, analyze-responses.js) that calls external MCP servers via HTTP, not an MCP server itself. Tool definitions are hardcoded in createMcpFunctionDeclaration() with no dynamic registration or schema validation. Error handling is present but minimal (basic HTTP status checks, no recovery guidance). No tool annotations (readOnlyHint, destructiveHint, idempotentHint). This scores as a test client, not a server.
Get weather forecast data for agricultural planning
Get weather forecast data for agricultural planning
No output schema documented. Tools return weather data but LLMs cannot plan downstream calls without knowing the response structure (fields, types, pagination).
Identical generic descriptions (54 chars) for both tools. LLMs cannot distinguish when to call GAP vs AccuWeather. Descriptions must explain the difference, data source, and when to prefer one.
Parameter descriptions are minimal (e.g., 'Latitude coordinate (e.g., -1.2864 for Nairobi)'). No guidance on valid ranges, precision, or when coordinates are required vs optional. Rubric baseline: 72 chars avg, these are ~50 chars.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 36 | 2026-07-28+ | v2 |
No error handling guidance. executeMcpToolCall() throws generic errors ('MCP server error', 'MCP error') with no recovery hints. LLMs cannot self-correct or retry intelligently.
This is a test client (batch-test-sdk.js), not an MCP server. It calls external MCP servers via HTTP POST to 'tools/call' endpoint. No MCP protocol implementation (no initialize, no tool registration, no resource/prompt handlers). Cannot be deployed as a standalone MCP server.