TokLedger provides 4 read-only tools with clear, domain-specific descriptions (avg 156 chars). All tools have proper JSON schemas with typed parameters. However, parameter descriptions are sparse or missing entirely (e.g., 'limit' in tokledger_recent lacks guidance on valid range or typical use). No tool annotations (readOnlyHint, idempotentHint) despite all being read-only. Error handling is minimal, no recovery guidance or actionable error messages visible in the MCP layer. Tool naming is verb-noun compliant but generic ('stats', 'recent', 'breakdown', 'savings'), could be more specific to the domain (e.g., 'get_inference_stats', 'list_recent_inferences'). Output schemas are not documented in the tool definitions.

• Add parameter descriptions to all inputs. For 'limit' in tokledger_recent, specify: 'Number of records to return (1-200, default 20). Larger limits may increase response time.'
• Document output schemas for each tool. E.g., tokledger_stats returns {request_count: int, prompt_tokens: int, completion_tokens: int, avg_ttft_ms: float, energy_kwh: float, local_cost_usd: float, cloud_cost_usd: float, savings_percent: float}.
• Add tool annotations: set readOnlyHint=true and idempotentHint=true on all four tools to signal safety to agents.
• Rename tools for clarity: tokledger_stats → get_inference_stats, tokledger_recent → list_recent_inferences, tokledger_model_breakdown → get_model_breakdown, tokledger_cloud_savings → get_cloud_savings.
• Implement error handling in the MCP server: return structured errors with actionable guidance (e.g., 'No ledger data available. Run inference through the proxy first.' or 'Invalid limit: must be 1-200.').