Most AI agents explore dbt projects the expensive way opening entire files and skimming thousands of irrelevant lines. We benchmarked this on a ~4,200 model dbt project: understanding a 2,455 line model costs 35,000 tokens via file reads. With structured retrieval via tree-sitter AST indexing, it's 3,000 tokens and adds critical business-context summaries instead of raw SQL to help your agent focus on what's important.
I'll share how I contributed a dbt context provider to jCodeMunch MCP that auto-enriches the code index with model descriptions, tags, and resolved {{ doc() }} column documentation from our project's schema.yml. I'll walk through benchmarked savings (64% token reduction, ~$11/session across 20 queries), the search_columns tool for cross-model column discovery (77% cheaper than Grep), and honest guidance on where structured retrieval wins versus where traditional tools still do.
The talk focuses on cost optimization, developer experience, and working with AI agents.