Lazy MCP gateway for DeepSeek Harness: one stable proxy tool instead of N tool schemas, servers connect on first use and idle out, metadata cached to disk.
Single tool 'mcp' acts as a gateway proxy rather than a direct tool. The tool definition is inferred from package.json and README context, not from explicit schema registration visible in source. Input schema shows 10 parameters with minimal type information (mostly strings and booleans). Description is 287 chars but generic, does not explain WHEN to use this tool vs calling servers directly, or what the LLM should expect from each operation mode (search vs describe vs tool invocation). No output schema documented. Tool violates single-responsibility principle by combining search, describe, and execution in one interface. Parameters lack descriptions explaining their purpose or constraints.
Gateway to every configured MCP server. Discover tools with { search }, inspect one with { describe }, call one with { tool, args }. Search and describe are answered from a local metadata cache and start nothing; a server starts only when one of its tools is actually called, then stops once idle.
Tool definition inferred from package.json, not explicitly registered in source code. No visible tool schema registration (e.g., via server.tool() or defineTool).
Input schema lacks type definitions for most parameters. 'search', 'describe', 'tool', 'args', 'server', 'connect', 'instructions' are all typed as strings or objects with minimal constraint. No enums, patterns, or ranges. Parameters 'regex', 'includeSchemas' are booleans but lack descriptions explaining their effect.
No output schema documented. LLM cannot predict what fields to expect from search, describe, or tool invocation results. Breaks downstream tool chaining and forces LLM to parse unstructured responses.
Tool combines four distinct operations (search, describe, tool invocation, server connection) into one interface. Violates single-responsibility principle. LLM must reason about which parameters to use for each intent, increasing error likelihood.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 37 | 2026-07-28+ | v2 |
Description does not explain WHEN to call this tool vs calling MCP servers directly, what each parameter mode does, or error recovery paths. Generic description wastes tokens and provides no selection guidance.