How Bonnard Builds Agent-Friendly MCPs
Exposing your data is easy. Designing an MCP that agents actually use well is the hard part. Right-sized tools, governed metrics, and token-efficient results.
Draft placeholder. Full article coming soon.
Exposing your data over MCP is the easy part. Designing one that an agent uses well is the hard part. This post will cover the techniques behind an agent-friendly data MCP.
What this will cover
- Discovery-first tools so the agent learns your schema instead of guessing
- A small set of purpose-built tools, not one tool per metric or a single SQL firehose
- A governed, structured query path as the default
- A narrow, audited escape hatch for the cases the structured path can't express
- Token-efficient, honest responses (pagination, completeness flags, measures over raw rows)
- Tool descriptions that double as the agent's instructions
Frequently asked questions
Why not just let the agent write SQL?
Text-to-SQL against raw tables is wrong often enough that you can't put it in front of a customer. A governed query path returns consistent, trustworthy numbers.
Does Bonnard support raw SQL at all?
Yes, as a narrow, audited escape hatch that still runs through the semantic layer, not a bypass around governance.
Ready to ship a customer-ready MCP?
Turn your semantic layer, dbt, or warehouse into a governed, per-customer MCP for your customers' agents.