Loading ToolBench
Minimal local RAG MCP server: semantic search over your markdown docs, embeddings via Ollama, nothing leaves your machine.
Single tool 'search_notes' has a well-structured schema with all parameters typed and described. Tool name is action-verb-based (search_*), appropriate for semantic search. Description is clear and contextual (98 chars). Parameters include query (required string), k (optional number with default context), exclude (optional array with detailed guidance), and searchAll (optional boolean with clear use case). Output is text-based search results with structured formatting. No critical gaps, but lacks explicit error handling guidance and output schema documentation.
Semantic search over a local markdown knowledge base (notes, design docs, decisions). Use it to recall past context and decisions.
Output schema not formally documented. Tool returns text content with formatted results, but no structured schema definition visible for downstream tool chaining or LLM parsing.
No error handling guidance. Tool lacks recovery instructions for edge cases (no results, embedding service unavailable, invalid query). LLM cannot determine next steps on failure.
Missing parameter constraints. 'k' parameter has no min/max bounds (default 8 mentioned but no range specified). LLM could pass absurd values like k=10000.
No tool annotations. Tool is read-only but lacks readOnlyHint annotation per current MCP spec (2026-07-28). Helps LLM understand safety profile.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |