Noisegate DP defines 5 tools with clear privacy-focused descriptions and reasonable parameter schemas. Tool names follow verb_noun convention (count, sum, avg, histogram, get_budget). Descriptions are domain-specific and explain the differential privacy guarantee. However, parameter schemas lack depth: 'filters' is typed as array but has no item schema or validation rules; column/group_by parameters lack enum constraints despite the code mentioning 'must be published in schema'. Output schemas are not documented. Error handling is mentioned in code comments but not surfaced in tool definitions. The server is STDIO-only, which caps protocol readiness at 50 regardless of definition quality.

• Add enum constraints to column and group_by parameters. Generate enums from DatasetPolicy.published_columns at server startup and embed in tool schemas so LLM can discover valid columns.
• Document filter structure in schema. Define a Filter object schema with operator (eq, lt, gt, in, etc.), column, and value fields. Expose this in the 'filters' array item schema.
• Document output schemas for all tools. Add a 'returns' section to each tool definition specifying field names, types, and meanings (e.g., count returns {value: int, margin: float, epsilon_spent: float}).
• Add error recovery guidance to tool descriptions. Mention that validation errors are returned as tool errors with rule details, and budget refusals are returned as kind='refusal' results (not errors), so LLM knows not to retry.
• Migrate from STDIO to Streamable HTTP transport. This enables remote deployment, hosted client support, and better observability. Use FastAPI's existing HTTP layer (api/main.py) as the transport backbone.