Offer a dry run for every change

How to do it

An agent that can preview a change makes fewer wrong ones, and can show the plan to its person first. Run the same code path up to the point of writing, inside a transaction you roll back, or with a flag that skips the write. Return the records that would change, the totals and any fees. Never send email, charge cards or call webhooks in a dry run. Mark the tool's other behaviour honestly, as Describe your MCP tools honestly says.

Examples

JavaScript example

export async function cancelBookings(args, user) {
  return db.tx(async tx => {
    const cancelled = await tx.any('UPDATE bookings SET status = $1 WHERE user_id = $2 AND day = $3 RETURNING id, fee', ['cancelled', user.id, args.day]);
    const plan = { cancelled: cancelled.map(row => row.id), fees: cancelled.reduce((sum, row) => sum + row.fee, 0) };
    if (args.dry_run) throw new RollbackWith(plan);
    await outbox.add(tx, 'cancellation-emails', plan.cancelled);
    return plan;
  }).catch(error => (error instanceof RollbackWith ? error.plan : Promise.reject(error)));
}

Python example

def cancel_bookings(args: dict, user) -> dict:
    with conn.transaction() as tx:
        rows = conn.execute("UPDATE bookings SET status = 'cancelled' WHERE user_id = %s AND day = %s RETURNING id, fee",
                            (user.id, args["day"])).fetchall()
        plan = {"cancelled": [row.id for row in rows], "fees": sum(row.fee for row in rows)}
        if args.get("dry_run"):
            tx.rollback()
            return plan
        outbox.add(conn, "cancellation-emails", plan["cancelled"])
        return plan

How to check

For each tool that changes data, a test calls it with dry_run: true and compares the database before and after. Nothing changed, no outbound message was queued, and the answer lists the planned changes.

Related rules

• WEB-196 Describe your MCP tools honestly in Website
• AGT-26 State the cost of every call before and after it runs in Agents as users