MCP server for semantic code search and Q&A using RAG
Server has 4 tools with complete schemas and descriptions. Naming follows verb_noun convention (index_, search_, ask_, list_). Descriptions are present (avg 140 chars) and explain WHAT each tool does. However, parameter descriptions lack actionable constraints (no enums, ranges, or format specs). Output schemas are not documented, LLMs cannot predict return structure. Error handling returns generic text responses without recovery guidance. No tool annotations (readOnlyHint/destructiveHint). Missing idempotency guarantees and pagination for list_repositories.
Ask a natural language question about an indexed codebase. Retrieves relevant code chunks and returns a grounded answer with source citations.
Crawl a local codebase, chunk files at function/class boundaries, generate embeddings, and store in vector DB for semantic search.
List all indexed repositories with metadata including chunk count, languages, and last indexed time.
Semantic similarity search over indexed code. Returns the most relevant code chunks for a natural language query.
Output schemas not documented. LLMs cannot predict return structure (fields, types, nesting). Forces agents to guess or fail on downstream tool chaining.
Parameter descriptions lack actionable constraints. 'limit' has no min/max (default 5 for search_code, but unbounded). 'include'/'exclude' glob patterns undocumented. LLMs cannot validate inputs before calling.
Error handling returns generic text ('Error: <message>') without recovery guidance. No categorization (retryable vs fatal). LLMs cannot decide whether to retry, ask user, or abort.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). index_repository is destructive (writes to vector DB) but unmarked. Agents cannot reason about side effects or retry safety.
list_repositories lacks pagination (no limit, offset, or cursor). If repo count grows, response bloats context window. No total count returned.