Server has 15 tools with complete input schemas and descriptions. Naming follows verb_noun convention (qllm_list_providers, qllm_chat, qllm_compare). Descriptions are present but vary in quality (50-250 chars, within baseline 194 avg). All parameters have type definitions and descriptions. However, output schemas are not documented, responses are returned as JSON strings without formal schema declarations. Error handling returns JSON with error fields but lacks recovery guidance. Tool annotations present (readOnlyHint, destructiveHint, idempotentHint) for most tools. No pagination support documented despite tools returning lists. Security: no secrets exposed as parameters (good), but no audit logging or permission gates visible.

• Document output schemas for all tools. Define response object structure with typed fields (e.g., 'Returns: {status: string, providers: [{id: string, protocol: string, enabled: boolean, ...}]}'). This enables LLMs to plan downstream tool calls.
• Add pagination to list tools: include limit (default 20, max 100), offset/cursor, and total_count in responses. Update qllm_list_providers, qllm_list_models, qllm_claude_sessions descriptions to document pagination behavior.
• Enhance error responses with recovery guidance. Replace _error_response() to return {status: 'error', error: string, error_type: string, recovery_hint: string, retryable: boolean}. Example: 'No enabled provider found. Try calling qllm_list_providers(require_secret=true) to see available options.'
• Clarify qllm_claude_capabilities purpose in description. Add 1-2 sentence explanation of when to call it (e.g., 'Call this to discover Claude Code feature matrix before invoking qllm_claude_run'). Expand from 50 to 100+ chars.
• Restrict extra_body parameter. Document allowed keys per provider (e.g., 'For OpenAI: top_k, frequency_penalty, presence_penalty. For Anthropic: budget_tokens. Invalid keys are silently ignored by the provider.'). Consider splitting into provider-specific parameters or validating against a schema.