Loading ToolBench
Closed Redis semantic data-plane tools exposed via Model Context Protocol
Redis MCP server has 15 tools with basic schemas and minimal descriptions. All tools have input schemas with proper JSON structure (oneOf patterns for operation dispatch), but descriptions are extremely terse (8-50 chars, well below the 194-char baseline). No parameter descriptions exist, only operation names and field names are visible. Output schemas are completely undocumented. Error handling and recovery guidance are absent. Tool naming follows a pattern (redis_<datatype>[_write]) but lacks action verbs (get_, set_, delete_) and is overly generic. No tool annotations (readOnlyHint, destructiveHint) despite clear risk classifications. Composition is reasonable (separate read/write tools), but parameter documentation is critically deficient.
Redis hash data structure read operations
Redis hash data structure write operations
Redis key operations (read-only)
Redis key operations (write)
Redis list data structure read operations
Redis list data structure write operations
Redis observation and monitoring operations
All tool descriptions are under 20 characters (e.g., 'Redis hash data structure read operations' = 42 chars, but most are 8-50 chars). LLMs cannot determine when to select these tools or what they return.
No parameter descriptions. Fields like 'key', 'field', 'value', 'operation', 'index' have no explanatory text. LLMs cannot infer whether 'key' is a Redis key name, a hash key, or a system identifier. Baseline: 100% of A+ tools have parameter descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 45 | 2026-07-28+ | v2 |
Redis set data structure read operations
Redis set data structure write operations
Redis sorted set data structure read operations
Redis sorted set data structure write operations
Redis stream data structure read operations
Redis stream data structure write operations
Redis string data structure read operations
Redis string data structure write operations
No output schemas documented. Tools return results but LLMs have no specification of response structure, field types, or what data is available for downstream tool chaining. Baseline: 100% of A+ tools have documented return types.
Tool names lack action verbs. 'redis_hash' does not convey 'get' or 'read'. Baseline: 90% of A+ tools start with action verbs (get_, set_, delete_, list_, search_). Current names force LLMs to read descriptions to understand intent.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite explicit risk classifications (READ_ONLY, WRITE, DESTRUCTIVE). Agents cannot distinguish safe from dangerous operations without reading descriptions.