MCP Postgres Inspector

The problem

AI agents are great at writing SQL and bad at knowing what your database actually looks like. Half the "hallucinated column" bugs I hit on stream came from the agent guessing at the schema.

The fix: give the agent eyes

The Model Context Protocol lets you expose tools to an agent. This server exposes exactly two, both read-only:

Tool	What it returns
`list_tables`	Every table with its columns, types and indexes
`sample_rows`	Up to 5 rows from a table (PII columns masked)

src/server.ts

server.tool(
  "sample_rows",
  { table: z.string(), limit: z.number().max(5).default(5) },
  async ({ table, limit }) => {
    // Whitelist the identifier. Never interpolate raw agent input.
    assertKnownTable(table);
    const rows = await sql`SELECT * FROM ${sql(table)} LIMIT ${limit}`;
    return { content: [{ type: "text", text: JSON.stringify(mask(rows)) }] };
  },
);

Guardrails

It connects with a read-only Postgres role, so even a bad prompt can't DROP anything.
Table names are checked against the introspected schema before any query runs.
Output is truncated so a wide table can't flood the agent's context.

The agent got noticeably better at writing migrations as soon as it could check the schema itself.