Server has 4 tools with basic schemas and descriptions, but significant gaps limit production readiness. Tool names lack clear action verbs (llms_bridge_call_* is verbose; call_chatgpt, call_gemini, call_perplexity would be clearer). Descriptions are present but generic (194-char baseline; these average ~120 chars, borderline). Parameter descriptions exist but lack constraints, temperature accepts 0-2 but no enum/validation guidance; model defaults are vague ('gpt-5.1' does not exist in OpenAI's current API). No output schema documented. Error handling returns generic LLMBridgeResult with success flag but no recovery guidance. Tool llms_bridge_server_misconfiguration has empty input schema and vague description. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Composition is reasonable, each tool calls one LLM, but parameter naming inconsistency (web_search_context_size as string enum vs temperature as number) and missing chaining IDs (no model/usage info in structured response) reduce usability.

• Remove or redesign llms_bridge_server_misconfiguration. If it returns config status, document the output schema (e.g., {available_models: [string], configured_apis: [string]}). If it's diagnostic-only, consider moving to a resource instead of a tool.
• Add enum constraints to parameters. For 'model', list valid options per provider (gpt-4o, gpt-4-turbo for OpenAI; gemini-2.0-pro for Google; sonar for Perplexity). For 'temperature', use min/max bounds in schema. For 'web_search_context_size', declare as enum: ['low', 'medium', 'high'].
• Document output schema in tool descriptions. E.g., 'Returns {response: string, model: string, usage: {prompt_tokens: int, completion_tokens: int, total_tokens: int}}'. This enables agents to extract usage metrics and chain calls.
• Enhance error messages with recovery guidance. Instead of 'Failed to call LLM: <error>', return 'API key missing or invalid. Ensure OPENAI_API_KEY is set in environment.' or 'Rate limit exceeded. Retry in 60 seconds.' Categorize as retryable vs user-fixable.
• Simplify tool names and add tool annotations. Rename to 'call_openai', 'call_gemini', 'call_perplexity' (drop prefix). Add readOnlyHint=true to all three (they do not modify state). This reduces cognitive load and clarifies intent.