Building Analytics Agents with DuckDB and MotherDuck (motherduck.com)

MotherDuck released a practical guide for building production-ready analytics agents on top of DuckDB that let users query data with natural language instead of writing SQL or dashboards. It walks teams through choosing an agent interface (Claude/OpenAI SDKs or Claude Desktop + MCP), trade-offs between generated SQL (flexible) and parameterized query templates (safer), and how to embed a DuckDB-tailored system prompt so the LLM knows syntax, exploration patterns, and performance best practices. The guide is significant because it operationalizes end-to-end agent design—schema discovery, secure access, query validation, and sandboxing—so organizations can deploy reliable, auditable data-facing assistants. Key technical details: recommended connectivity uses DuckDB’s MotherDuck URI (e.g., duckdb.connect('md:my_database?motherduck_token=<token>', read_only=True)); use MCP server or custom function tools (query_motherduck) to execute SQL; include function_docs.jsonl snippets (top 5 relevant functions) in context to help with specialized operations. Schema discovery helpers include __MD_FILTER_TABLES(), duckdb_tables(), and duckdb_columns(); performance tips cover QUALIFY, arg_max, early filters, and avoiding functions in WHERE. Security/ops advice covers read-only tokens, per-tenant service accounts, and zero-copy clones for writable sandboxes. For safety, validate queries with CALL try_bind(...) before presenting or executing them.