This server provides three well-defined conversion tools with clear names, good descriptions, and complete input schemas. All tools follow the verb_noun naming pattern (convert_unit, get_conversion_factor, list_units). Descriptions are substantive (150-200 chars), exceeding the 10-char minimum and within the 10-1024 baseline. Input parameters have types and descriptions. However, output schemas are not formally documented in the code, they are returned as JSON strings without explicit schema declarations. Error handling is good (validates inputs, returns helpful messages with available options), but does not classify errors as retryable/user-fixable/fatal. The server lacks tool annotations (readOnlyHint, destructiveHint, idempotentHint) that would signal to agents these are safe, read-only operations. Resources are well-structured but ancillary to the core tools.

• Add formal output schemas to each tool definition. Document the structure of the JSON response (e.g., convert_unit returns {input: {value, unit}, output: {value, unit}, formula: string}). Use MCP's output schema field in tool definitions.
• Annotate all three tools with readOnlyHint: true and idempotentHint: true. This signals to agents that these calls are safe to retry and do not have side effects.
• Add enum constraints to the category parameter in list_units. Change the parameter definition to specify enum: ['weight', 'length', 'volume'] so LLMs cannot pass invalid categories.
• Enhance error messages to classify errors. When validation fails, return a structured error object with type (e.g., 'invalid_unit', 'incompatible_categories') so agents know whether to retry, ask the user, or give up.
• Consider adding a discovery tool or endpoint that lists supported units and categories upfront. The current list_units is good, but agents might benefit from a 'describe_conversion_system' tool that returns the entire taxonomy in one call.