Introducing Google’s MCP Toolbox for Databases and how agents use it with ADK to run BigQuery tools like schema inspection and SQL execution to automate analytics queries
It’s Monday morning again — the first day of week 2 — and Claire’s name is already at the top of your inbox.The subject line reads, “One more question, sorry.”Ah, well, Mondays.Let’s open the email.Which five accounts have the sharpest usage drops right now?You know you’ll keep getting this question. Last week you’d open your editor and start writing a helper like list_sharpest_drops. This week you skip hand-writing that function and attach an MCP server instead — fulfilling last week’s promise to automate query execution for the agent.A few small helper functions you might have written by hand look like:
def get_customer(customer_id): """Return customer metadata for a given customer_id.""" passdef list_sharpest_drops(n=5, lookback_days=14): """Return the top `n` accounts with the largest usage drops over `lookback_days`.""" pass
Authentication doesn’t change: your existing Application Default Credentials still cover BigQuery. The same .env file keeps your Google and model API keys and project ID:
Next comes the server: Google’s MCP Toolbox for Databases — an open-source MCP server built exactly for this job. It connects to databases and publishes ready-made tools that agents call directly, so you no longer have to maintain custom query-running logic.
The Toolbox ships as a single self-contained binary — no installer or runtime to manage. Running it is a two-step process: download the binary, then make it executable.
You never manually run a long-lived server process. The ADK launches the Toolbox as a subprocess and communicates via STDIO: ADK writes MCP requests to the Toolbox’s stdin and reads MCP responses from its stdout. Start the prebuilt BigQuery plugin like this:
./toolbox --prebuilt bigquery --stdio
Using --stdio makes the Toolbox exchange MCP messages over standard input and output instead of a network socket. This lets the ADK spawn, supervise, and terminate the Toolbox as a local subprocess.
Toolbox publishes nine ready-made BigQuery tools; two are especially relevant for this scenario:
Tool name
Purpose
execute_sql
Runs a SQL query generated by the model and returns the query results.
get_table_info
Returns a table schema and metadata so models can craft syntactically correct queries for your dataset.
These tools let the agent request schema information and execute SQL directly against BigQuery without you writing or maintaining the plumbing.
With the MCP Toolbox running as a subprocess, the agent can focus on asking higher-level questions — for example, “Which five accounts have the sharpest usage drops?” — and rely on get_table_info to inspect schemas and execute_sql to run the resulting queries. This reduces maintenance, centralizes database access, and improves safety by keeping SQL execution within the managed MCP server layer.Links and references
