Three tools with partial schema definitions and minimal descriptions. Tool names follow verb_noun convention (execute_query, get_tables, describe_table), naming is acceptable. However, descriptions lack depth and action context. Input schemas are present and properly typed, but output schemas are completely undocumented, LLMs cannot reason about what these tools return. No error handling guidance, no parameter constraints (e.g., enums, ranges), and no indication of which operations are idempotent or have side effects. The execute_query tool accepts 'params' as an array of strings with zero documentation on how parameters are bound or escaped, a serious security gap and a source of LLM confusion. get_tables and describe_table descriptions are generic ('Get list of tables', 'Get column information') and do not explain when to use each or what structure they reveal. Overall, this is a typical C-grade community server with incomplete tooling.
execute_query params parameter accepts array of strings with zero explanation of binding mechanism, parameter names, or SQL injection protection, creates security risk and LLM confusion
No error handling guidance, responses may return errors but LLMs have no indication of what errors mean, whether to retry, or how to recover (e.g., 'table not found' should suggest searching available tables)
Recommendations
Document output schemas for all three tools. Example: execute_query should return {rows: [{column_name: any},...], rowCount: number, executionTimeMs: number}. get_tables should return {tables: [{name: string, owner: string, type: string},...], total: number}. describe_table should return {columns: [{name: string, dataType: string, nullable: boolean, primaryKey: boolean, defaultValue: string},...]}.
Rewrite execute_query description to explain parameterized queries: 'Execute a SQL query with optional bind parameters. Parameters are bound safely using Oracle bind variable syntax, pass parameter values in the params array in the same order they appear in the query (e.g., query="SELECT * FROM users WHERE id = :1 AND status = :2", params=[123, "active"]). This prevents SQL injection. Read-only queries are recommended; destructive queries (UPDATE, DELETE, DROP) may be restricted by server configuration.'
Add parameter constraints: specify that query must be non-empty and less than 10000 characters; params array should contain at most 100 items; owner (if provided) must be 1-30 alphanumeric characters or uppercase schema name.
Enhance get_tables description: 'List all tables accessible to the current database user, optionally filtered by schema/owner. Returns table names, ownership, and type (TABLE, VIEW, etc.). Useful for discovery before calling describe_table to inspect column structure. Note: result is capped at 500 tables; use owner filter to narrow results if needed.'
Enhance describe_table description: 'Retrieve detailed metadata for a specific table: column names, data types, nullability, primary key status, and default values. Use this after get_tables() to inspect table structure before writing queries. Returns up to 1000 columns per table.'
No pagination documented, tools returning lists (get_tables, describe_table columns) may exceed context window; no mention of limits, cursors, or offset/limit parameters
No security boundaries, execute_query accepts arbitrary SQL with no indication of read-only enforcement, permissions, or timeouts; no clarification whether destructive queries (UPDATE, DELETE, DROP) are allowed
Parameter owner in get_tables and describe_table lacks format constraints, is it a schema name, username, or role name? What happens if it does not exist?
get_tablesdescribe_table
Add error guidance: 'If execute_query returns "table not found", call get_tables() to list available tables. If a column name is invalid, call describe_table() first to verify spelling and type.'
Implement and document result limits: get_tables returns max 500 rows, describe_table returns max 1000 columns. State these in descriptions so LLM knows when to filter or paginate.
Add a 'dry_run' boolean parameter to execute_query (default false) to let agents preview query results without side effects, useful for validating queries before they execute against production.
Document the owner parameter clearly: 'Oracle schema name (e.g., "HR", "SYSTEM"). If omitted, defaults to the current user\'s schema. Case-sensitive in most Oracle configurations, verify exact case with get_tables().'
Add timeout and row limit to execute_query description: 'Queries timeout after 30 seconds. Result sets are capped at 10000 rows to prevent context explosion, use LIMIT clause for larger queries.'