Bamboo MCP Gateway has 5 tools with explicit schemas and descriptions visible in src/mcp-server.ts. However, the server suffers from critical gaps in parameter descriptions, output schema documentation, error handling guidance, and security considerations. The tools follow a namespace.verb naming pattern (ads.get_*, pg.query, pg.get_tables) which is reasonable, but parameter descriptions are minimal or absent. Most critically: pg.query accepts arbitrary SQL without input validation documentation, poses SQL injection risk, and lacks any error recovery guidance. Output schemas are entirely undocumented, the code registers tools but never documents what the tools return, forcing LLMs to guess structure. No tool annotations (readOnlyHint/destructiveHint) are present despite clear distinctions between read and write operations. Parameters like 'params' in pg.query lack type constraints (array items should not all be 'string'). No confirmation patterns for destructive operations (pg.query with DELETE/DROP). The implementation shows the basic MCP scaffolding (tools/list, tools/call, resources/list, resources/read) but definition quality is below median community standard.

• Expand all tool descriptions to 100-250 characters. For ads.get_campaigns: 'Retrieve Meta Ads campaigns from a specific ad account. Required for discovering campaign IDs needed by get_adsets. Returns campaign name, ID, status, and spend totals. Requires account_id (e.g., "act_123456"). Limit defaults to 25, use offset for pagination.'
• Document complete output schema for each tool. E.g., ads.get_campaigns returns {campaigns: [{id: string, name: string, status: string, spend: number}], total: number, next_cursor?: string}
• Add strict input validation and error messages to pg.query. Return 'Invalid SQL: detected DROP/TRUNCATE. Use read-only queries or contact admin.' for destructive ops. Return 'SQL Error: <detail>. Check syntax and parameter types.' for parse failures.
• Implement parameterized query framework. Change pg.query to enforce prepared statements with separate query + typed params. Update schema: params: {type: 'object', properties: {param_name: {type: 'string|number|boolean'}}} to support multiple types.
• Add tool annotations to input schemas. For ads.get_campaigns and all reads: append '"readOnlyHint": true'. For pg.query: append '"destructiveHint": true, "idempotentHint": false'. (Note: annotations property structure depends on MCP spec version, verify current format.)
• Add pagination support. For ads tools: add offset/cursor parameter and return total_count and next_cursor. Document: 'If results are truncated, use offset + limit or next_cursor for subsequent calls.'
• Implement dry-run confirmation for pg.query. Add optional dry_run boolean param. If true and query is destructive, return 'Would delete X rows. Set dry_run: false to execute.' This prevents accidental data loss.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry