Most teams connecting AI agents to their data warehouse start with text-to-SQL. The agent generates SQL from natural language, runs it against the warehouse, and returns results. It works until it doesn't: hallucinated JOINs, inconsistent aggregations, no access control, no audit trail.
There's a better approach. Define your business metrics in a semantic layer, expose them via MCP (Model Context Protocol), and let any AI agent query governed definitions instead of raw tables. Then add one tool so the agent can chart the result in Claude or ChatGPT. This tutorial shows how to set it up in under 30 minutes.
The agent sees column names but not business logic. It doesn't know that your company excludes refunds from revenue. It doesn't know that status = 'completed'
means something different in orders
than in subscriptions
. It doesn't know that marketing and finance defined "active user" differently three years ago and never reconciled.
So the agent writes plausible SQL and returns plausible numbers. Ask the same question twice with different phrasing and you get different answers. Ask two different agents and you get two different numbers. Neither matches the number your finance team reports.
Beyond consistency, there's no row-level security. No multi-tenancy. No audit trail showing which agent queried what, when, and for whom. In production, with real customers, that's a non-starter.
Text-to-SQL gives you speed. It doesn't give you trust.
Instead of letting agents write arbitrary SQL, define your metrics once in YAML: cubes, measures, dimensions, access rules. Then expose those definitions via MCP so agents query governed metrics, not raw tables.
The difference: every agent gets the same answer because the metric definition is fixed. total_revenue
isn't a column the agent interprets. It's a pre-defined calculation with agreed-upon filters and aggregations. When your finance team updates the revenue definition to exclude trial conversions, that change propagates to every consumer instantly. No agent retrained. No dashboard patched. One diff in your schema repo.
This architecture also decouples the query interface from the warehouse dialect. Swap BigQuery for Snowflake and your agents don't notice. The semantic layer abstracts the SQL generation, so consumers stay stable while infrastructure evolves underneath.
| Approach | Governance | Consistency | Multi-tenant | Access Control | Setup Time |
|---|---|---|---|---|---|
| Direct SQL | None | Varies by query | Manual | Manual | Minutes |
| Text-to-SQL | None | Varies by prompt | Manual | Manual | Hours |
| Semantic Layer via MCP | Full | Guaranteed | Built-in | Row-level | Under 30 min |
The semantic layer is the control plane between your warehouse and every consumer, whether that's a human analyst, a React component, or an AI agent.
Stand up a semantic layer that connects to your warehouse and exposes metrics over MCP. Open-source engines like Cube connect to BigQuery, Snowflake, Redshift, Databricks, PostgreSQL (including Supabase, Neon, and RDS), and DuckDB (including MotherDuck). You configure your warehouse connection in a config file or pass credentials via environment variables.
If you want to explore without connecting your own warehouse, point the semantic layer at a sample PostgreSQL database with orders, customers, and products data so you can follow along with the rest of this tutorial. A few thousand rows across three tables is enough to test aggregations, filters, and multi-dimensional queries.
Create a cube that maps to a table in your warehouse and defines the metrics your agents will query.
cubes:
- name: orders
sql_table: public.orders
measures:
- name: total_revenue
sql: amount
type: sum
- name: count
type: count
dimensions:
- name: status
sql: status
type: string
- name: created_at
sql: created_at
type: time
Measures are the numbers you aggregate: sums, counts, averages. Dimensions are the columns you filter and group by: status, date, category. One definition. Every tool, dashboard, and AI agent that queries total_revenue
gets the same number.
You can also define pre_aggregations
in the same file to cache expensive computations. For example, a daily rollup of total_revenue
by status
can cut query times from seconds to single-digit milliseconds. The semantic layer rebuilds these rollups on a configurable schedule and invalidates stale caches automatically.
Schemas are version-controlled alongside your application code. Review metric changes in pull requests. Roll back a bad definition with git revert
. Your data contracts get the same CI/CD workflow as your product.
To control what's exposed to specific consumers, define a view:
views:
- name: order_metrics
cubes:
- join_path: orders
includes:
- total_revenue
- count
- status
- created_at
Views act as a curated interface. Your agents see order_metrics
with four fields instead of navigating the full schema.
Deploy your schema so the semantic layer serves it over an MCP endpoint, then add the MCP server URL to your client's config:
{
"mcpServers": {
"semantic-layer": {
"type": "http",
"url": "https://your-semantic-layer.example.com/mcp"
}
}
}
Paste this into your MCP client's config file and restart. For Claude Desktop, that's ~/Library/Application Support/Claude/claude_desktop_config.json
. For Cursor, it's .cursor/mcp.json
in your project root. Claude Code reads from .mcp.json
in your project directory.
The MCP server handles tool discovery, schema introspection, and query execution over HTTP. Your agent sees available metrics the same way it sees any other MCP tool. No custom integration code required.
For customer-facing use cases, scope each connection to a specific tenant's data with row-level security defined in the schema, so a customer's agent sees only that customer's rows.
Once connected, your AI agent can discover and query your metrics using natural language. Behind the scenes, the agent calls MCP tools.
Ask: "What's our total revenue this quarter?"
The agent calls explore_schema
to discover available metrics, then calls query
with the right measures and time filters. The response comes back as structured data, not raw SQL results.
Ask: "Break down order count by status for the last 30 days."
Same flow. The agent uses query
with count
as the measure, status
as the dimension, and a date filter on created_at
.
A semantic layer typically exposes a handful of MCP tools:
explore_schema
query
sql_query
describe_field
Every query runs through the semantic layer. The agent never touches raw tables. Read more about the architecture in our agentic analytics guide.
A governed query returns rows. The agent usually needs to show the user a chart, not a wall of numbers. Leaving the model to draw HTML puts the most important part of the answer in its hands to improvise.
@bonnard/mcp-charts fixes that. It adds a
visualize
tool to your MCP server in a few lines:
npm install @bonnard/mcp-charts
js
import { addCharts } from "@bonnard/mcp-charts";
// your data, your connection. Bonnard never touches the database
addCharts(server, { runSql });
The agent calls visualize
with a query, your runSql
returns the rows, and Bonnard infers the chart from the typed result, then renders an interactive widget in Claude or ChatGPT. It renders line, bar, area, pie, scatter, funnel, waterfall, and table, with bar variants for stacked, grouped, horizontal, and 100% stacked. The chart comes from your query result, not from tokens the model invents. Same data, same chart, every time. Full walkthrough: MCP Charts.
You started with a warehouse and an AI agent that writes its own SQL. Now you have governed metrics defined in YAML, exposed via MCP, queryable from any AI tool your team or customers use. Setup took under 30 minutes.
The shift from raw SQL to governed metrics pays off immediately: consistent numbers across every consumer, row-level access control per tenant, and a full audit trail for every query. As your team adds more agents and surfaces, the semantic layer scales with you. No duplicate logic. No drift between dashboards and AI answers.
Give the agent a way to chart what it queries: npm install @bonnard/mcp-charts
, call addCharts(server, { runSql })
, and interactive charts render in Claude and ChatGPT. The repo is on GitHub. Read the MCP Charts guide to go deeper.
visualize
tool so agents chart query results in Claude and ChatGPTNo. A semantic layer with MCP support works with any MCP-compatible AI agent out of the box. Your agent discovers available metrics through the MCP protocol at runtime. There is no fine-tuning, prompt engineering, or model modification required.
A semantic layer connects to BigQuery, Snowflake, Redshift, Databricks, PostgreSQL (including Supabase, Neon, and RDS), and DuckDB (including MotherDuck). For charting the agent's results, @bonnard/mcp-charts
ships native adapters for Postgres, BigQuery, Snowflake, Databricks, and DuckDB, or you pass your own runSql
. Bonnard never touches your database.
Text-to-SQL lets an AI agent generate arbitrary SQL from natural language. It has no governance, no consistency guarantees, and no access control. A semantic layer defines metrics once in YAML and exposes them as governed APIs. Every consumer gets the same answer because the calculation is fixed, not interpreted per query. And when the agent charts the result, @bonnard/mcp-charts
renders from the query result, not from tokens the model invents.
Install @bonnard/mcp-charts and call
addCharts(server, { runSql })
on your MCP server. That registers a visualize
tool. The agent calls it with a query, your callback returns the rows, and an interactive chart renders in Claude or ChatGPT. No frontend code.Under 30 minutes for most teams. Stand up a semantic layer, connect your warehouse, define a few metrics in YAML, and deploy it over MCP. Add charts with npm install @bonnard/mcp-charts
and one addCharts
call. The tutorial above walks through each step with working code examples.