The server has 3 clearly defined tools with explicit registration via @mcp.tool() decorator. All tools have descriptions and input schemas visible in the source code. However, there are significant gaps in parameter documentation, output schema definition, and error handling guidance. Naming is verb-noun compliant (predict_, estimate_, get_), but descriptions vary in quality and some parameters lack descriptions entirely. The tools accept blockchain-specific parameters that are reasonably well-constrained (chain_id, confidence 0-100, gas_limit), but output schemas are not formally documented, responses are string-based Markdown tables rather than structured objects. Error handling exists but is minimal and does not guide LLM recovery. The server is neither exceptional nor failing, landing solidly in the 'fair' range with room for improvement.

• Convert all tool responses from plain Markdown strings to structured JSON objects. For example, predict_gas_price should return {"chain_id": 1, "system": "ethereum", "network": "mainnet", "base_fee_per_gas": 45.2, "predictions": [{"confidence": 99, "price": 50.1, ...}]} instead of a formatted string. This enables LLMs to extract and chain values reliably.
• Add pagination to get_supported_chains and predict_gas_price. Accept optional 'limit' (default 20, max 100) and 'offset' (default 0) parameters, and return {"items": [...], "total": 500, "limit": 20, "offset": 0} to support large result sets without context window exhaustion.
• Enhance error messages with recovery guidance. Instead of 'Error: Confidence must be between 0 and 100', return 'Invalid confidence: got 150. Must be 0-100. Please retry with a valid value.' Similarly for gas_limit and other constraints.
• Remove example values from parameter descriptions. Change 'The ID of the blockchain network (e.g., 1 for Ethereum Mainnet)' to 'The blockchain chain ID (e.g., Ethereum mainnet uses 1, Polygon uses 137). Call get_supported_chains() to list available chain IDs.' This guides discovery without forcing example reuse.
• Add tool annotations (readOnlyHint) to all three tools. They are read-only and safe to call repeatedly, mark them explicitly in the MCP schema so clients know they are safe for agentic exploration.
• Document output field names and types explicitly. Create a formal schema (even as a comment) for each tool's response, e.g., 'Returns object {"base_fee_gwei": float, "predictions": [{"confidence": int, "price_gwei": float, "max_fee_gwei": float}]}'.