SQLHarness

Ask the question.
Govern the query.

Your existing AI model proposes SQL. A shared database authority validates the query and runs it through read-only, bounded SQLite execution.

Try an example ↓Add samples to VS Code ↗

No sign-in. No extra API key. No personal uploads.

One prompt. Real tool output.

5 governed database tools

A synthetic shop with three customers and four orders. Total paid revenue: $1,220. Choose an example to make a real MCP call. This page shows the evidence; your AI client supplies the conversational answer.

EXAMPLE PROMPTS
Ready when you areEXAMPLE DATA ONLY
Choose a prompt on the left.No model call or API key is needed.

Connect once. Ask naturally.

Add this Streamable HTTP server in an MCP-compatible AI client. It exposes the same public example data as the demo above. Your normal AI account and usage apply.

https://site--codearchitect--74wvcb9qzxr6.code.run/mcp/sqlharness/server
Add to VS Code ↗
Configuration for VS Code and compatible clients
{
  "servers": {
    "sqlharness": {
      "type": "http",
      "url": "https://site--codearchitect--74wvcb9qzxr6.code.run/mcp/sqlharness/server"
    }
  }
}

Other clients may use a different configuration format. Choose HTTP / Streamable HTTP and paste the endpoint above. No authorization header is required.

Then ask your model

Bring the tools to your own work.

SQLite databases in a folder you explicitly select. SQLGlot policy checks, execution timeout, and row limits.

uvx --from https://site--codearchitect--74wvcb9qzxr6.code.run/mcp/install/sqlharness_mcp-0.2.1-py3-none-any.whl sqlharness-mcp --database-root /absolute/path/to/databases

Requires Python 3.11+ · uv / uvx.

Add this command as a local stdio MCP server in your AI app. Place a database at databases/shop/shop.sqlite and use shop as its ID. Your model writes SQL; the server does not call another model.

Download the versioned install package ↗
What “local” means for privacy

These MCP tools do not save request or response content, enable telemetry, or call a model provider. They return selected evidence to your AI host. That host may retain it or send it to its model provider under its own policy.

SQLHarness opens databases in the folder you select in read-only mode. Query inputs, result rows, and validation receipts are not saved by the MCP server. Your AI client generates SQL; the server does not call another model.