What it does
BigQuery Metric Narrator reads a read-only (SELECT-only) metric query result from BigQuery or a warehouse — or accepts a provided query-result export — and turns it into a deterministic analysis: period-over-period change percentage, trend, severity, breakdown share, outlier detection, and a confidence score. An LLM then writes a grounded narrative and decision-support recommendations; every number is code-derived and the LLM never invents or alters figures. Gated: ships with mock fixtures and a provided-export path so it works with no live network; live BigQuery verification is pending a sandbox key. Guardrails: read-only (DML/DDL rejected), credential masking, prompt-injection defang, query row/byte/timeout limits, never a production dataset. Built for SaaS analytics and data teams who want automated warehouse metric narration instead of manually eyeballing dashboards.
Example prompts
- Analyze this month's active users and MRR from our BigQuery export
- What changed period-over-period in these metrics and why?
- Validate that this SQL query is read-only before running it
- Break down MRR change by customer segment and flag outliers
Tools (7)
Tools the agent exposes — your AI client calls them automatically when it needs them.
- run_full — Run the agent end-to-end: collect a read-only metric query result, compute deterministic deltas/trends/breakdowns/outliers, synthesize narrative + recommendations.
- analyze_metrics — Given raw collected data, run the pure deterministic analysis (no LLM): change %, trend, severity, breakdown share, outliers and confidence.
- compute_change — Compute the deterministic period-over-period change percentage and trend for a single metric value/prev pair.
- validate_query — Validate that a query is read-only (SELECT/WITH only) before it could run against the warehouse.
- list_capabilities — List the agent's static capabilities: tools, thresholds, guardrails and credential slots.
- discover_intent — Understand your goal and co-design the exact input via clarifying questions before running.
- plan_inputs — Plan/brainstorm the inputs for a tool: returns questions, schema and a ready-to-edit example.
What you'll need to connect
This agent will ask you for the following. You enter them when you connect — they're encrypted and never shared with the creator.
- BigQuery Project ID (gated) · optionalGATED. GCP project id used for read-only BigQuery queries. Leave blank to use provided export rows / mock fixtures.In Google Analytics, find your property ID under Admin → Property Settings. Then create a service-account JSON key in Google Cloud (IAM & Admin → Service Accounts → Keys → Add key → JSON) and give that service account Viewer access to the GA4 property.Paste the value as a single line.Only sent to: bigquery.googleapis.com, oauth2.googleapis.com
- Service account JSON · optionalIn Google Analytics, find your property ID under Admin → Property Settings. Then create a service-account JSON key in Google Cloud (IAM & Admin → Service Accounts → Keys → Add key → JSON) and give that service account Viewer access to the GA4 property.Paste the full JSON of your service-account key file (not a file path). FindAgent stores it securely and gives the agent a file path to it.Only sent to: bigquery.googleapis.com, oauth2.googleapis.com
- Anthropic API Key · optionalOptional. Leave blank on a sampling-capable client to use the host model with no key.Create a key on the Anthropic Console API keys page (console.anthropic.com → API keys).Paste the value as a single line.Only sent to: api.anthropic.com
How you're protected
FindAgent runs these safety checks on every agent automatically. They're always on and can't be turned off.
- Prompt-injection scanning
Every request is checked for known prompt-injection and jailbreak attempts before the agent runs. This is always on.
- Secret-leak scanning
Every response is scanned for leaked API keys, tokens, and other secrets before it reaches you. This is always on.
