An MCP server that indexes and searches a curated subset of tools using regex full-text search and BM25-ranked search, with tool catalog management via a web-based schema editor
Master-MCP provides two well-implemented search tools with complete schemas and reasonable descriptions. Both tools have proper input schemas with typed parameters and descriptions. However, there are notable gaps in output schema documentation, parameter descriptions lack specific constraints and examples, and error handling guidance is absent. The tools are READ_ONLY which is good for safety, but descriptions do not explicitly state this or provide recovery guidance for edge cases. Parameter descriptions are adequate but could be more detailed (current median ~50 chars vs. production baseline of 72 chars). Output schemas are not documented, the code mentions returning 'tool name/description/inputSchema' but the actual response structure is not formally specified. The BM25 implementation shows solid engineering, but the tool interface does not follow some key patterns around error classification and actionable error messages.
BM25 ranked search over curated Master-MCP tools (schemas/search_tools.json). Returns tool name/description/inputSchema; in CodeExec mode returns server/function/parameters.
Regex full-text search over curated Master-MCP tools (schemas/search_tools.json). Returns tool name/description/inputSchema; in CodeExec mode returns server/function/parameters.
Output schemas not documented. The inputSchema is complete with types and descriptions, but responses return tool name/description/inputSchema without a formal outputSchema specification. LLMs cannot plan downstream tool calls without knowing the response structure.
Parameter descriptions lack constraints and formatting hints. The 'pattern' parameter in search_regex_tools says 'Regex pattern' but does not specify regex flavor, length limits, or invalid patterns. The 'flags' parameter lists valid values but as free-form text in the description instead of an enum. LLMs will hallucinate invalid flag combinations.
No error handling or recovery guidance. If the search_tools.json file is missing, empty, or malformed, the tools will return empty results or crash silently. The descriptions do not explain what happens on no matches, or guide the LLM to try alternative searches or ask the user for clarification.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 53 | 2025-06-18+ | v2 |
| 2026-03-09 | D | 52 | - | v1 |
Tool descriptions do not explicitly state idempotency or safety. Both tools are marked READ_ONLY, which is excellent for safety, but the descriptions do not say 'This tool is read-only' or 'No side effects'. LLMs reason better when safety properties are explicit.
The 'flags' parameter in search_regex_tools should be an enum, not a free-form string. Current description lists 'i (ignorecase), m (multiline), s (dotall)' as examples but does not prevent LLMs from passing invalid flag combinations like 'flags: "xyz" ' or using PCRE-specific flags.
CodeExec mode behavior not documented. The code checks for MASTER_MCP_CODEEXEC environment variable and changes response format (returning 'server/function/parameters' instead of tool name). This branching behavior is not mentioned in tool descriptions, making it impossible for LLMs or users to understand when and why output changes.