Run a reviewable investigation.

QueryBridge has a credential-free synthetic demo and optional real PostgreSQL, BigQuery, and Claude integrations. Run the local components on your machine.

Synthetic demo walkthrough

The screenshot shows the actual local application, not an online database session. In the procurement fixture, two receipts repeat one order’s quantity: the join sums to 250 units while the base orders total 150. An unmatched receipt and a separate one-unit inventory discrepancy provide additional investigation examples. All data is synthetic.

Run locally

Early-access distribution is arranged by email. Once you have the source, install Node.js 24 LTS and run from the repository root:

npm ci
npm run dev

Open the local workbench address printed by the launcher. The database API listens only on your machine. This public site does not connect to it or provide a hosted database demo.

1. Select “Procurement sandbox.” It is explicitly labelled synthetic.
2. Review the selected receipt join. Ask “Why does this JOIN duplicate quantities?”
3. The scripted assistant inspects the selected schema and creates pending checks.
4. Review the exact SQL, connection, parameters, scope, and byte limit. Approve each check.
5. Preview aggregate-only or redacted evidence. Approve the exact payload to produce the local scripted summary.
6. Open each supporting query and result, review uncertainty and the correction, and export Markdown.

Real connections

Configure PostgreSQL credentials only in the local environment. BigQuery uses Application Default Credentials, a billing project, dataset location, and an explicit maximumBytesBilled. Configure allowed relations and narrow the table scope in the workbench.

Use a dedicated database identity with only required metadata and SELECT access. The PostgreSQL read-only transaction supplements permissions. A LIMIT is a row bound, never a BigQuery cost cap.

Claude context

Set ANTHROPIC_API_KEY and ANTHROPIC_MODEL server-side. Starting a live investigation sends selected SQL and approved column metadata. Results require a separate exact-payload approval. The model can create proposals but cannot authorize execution.

Supported SQL

One SELECT statement, simple joins, non-recursive read-only CTEs, grouping, common comparisons, and explicitly allowed aggregate/scalar functions. Fully qualify physical tables. Writes, DDL, multiple statements, procedures, exports, table functions, external tables, views, recursive CTEs, and unsupported clauses are rejected.

Operational guides

The repository includes architecture, credential setup, deployment, domain/email, verification, and roadmap guides under docs/. Live BigQuery and Claude tests are optional and credential-gated.

Development stage

This is an MVP with a deliberately small SQL subset. It is not a general-purpose database security gateway, and bounded evidence does not prove complete reconciliation. SQL Server, Snowflake, and Databricks remain roadmap items.