Two tools with clear verb-noun naming (get_costs, get_projects) and complete input schemas. Descriptions are present but generic (10-15 words each). All parameters have type annotations and descriptions via Annotated. However, output schemas are undocumented, LLMs cannot predict return structure. No error handling guidance. Parameters lack constraints (e.g., no limit on group_by cardinality, no validation of datetime ranges). Tool descriptions do not explain WHEN to use each tool or what distinguishes them. Missing pagination documentation despite implementing pagination internally.

• Document output schemas explicitly. Add return type hints like 'Returns: list[{id: str, name: str, created_at: datetime, ...}]' to docstrings or use Pydantic models to auto-generate schemas.
• Expand tool descriptions to 50-200 chars with context. Example: 'get_costs: Fetch OpenAI organization spending for a date range, grouped by project or line item. Use this to analyze billing trends or audit usage. Returns costs in USD with timestamps.' Include WHEN to use it.
• Add error handling documentation. Document that 401 means invalid OPENAI_ADMIN_API_KEY (user-fixable), 429 means rate limited (retryable with backoff), 500 means API down (retryable). Return structured error responses with recovery hints.
• Expose pagination as tool parameters. Add optional 'limit' (1-180) and 'offset' or 'cursor' parameters so LLMs can control batch size and iterate. Document that results are capped at 180 per call.
• Add parameter constraints to descriptions. Example: 'group_by: List of grouping fields (max 2). Valid: project_id, line_item.' and 'start_time: UTC timestamp, must be within last 90 days.' Validate in code and return actionable errors.