SMB Sandbox MCP server

smb-sandbox

smb-sandbox is an open-source, fully fictional small business for testing AI agents: a realistic dataset generator, mock CRM, inbox, calendar, invoicing and phone systems exposed as MCP servers, and verified business tasks graded on the end state of the business, not on what the agent says it did.

Agents that look great in a demo often fall apart on a real company's messy data: a duplicate contact, a quote nobody followed up on, a payment that doesn't match any invoice. You can't safely test against real customers, and random fake rows don't have those problems. smb-sandbox gives you a believable company that does, with every problem labeled so you can measure whether your agent fixed it.

What's inside

Three starter industries: home services (HVAC and plumbing), dental clinic, and marketing agency. Each seed is a different company with the same kinds of problems.

Evaluate an agent in three commands

# 1. Start the sandbox business and record everything the agent does
npx @teamshift/sandbox-mcp --industry home-services --seed 42 --toolsets crm,inbox,calendar,invoicing,phone --state-out run.json

# 2. Give your agent the task prompt (Claude Code, Cursor, or your own agent connected to the server above)
npx @teamshift/smb-workflows show quote-follow-up --industry home-services --seed 42

# 3. Grade the end state
npx @teamshift/smb-workflows grade quote-follow-up --industry home-services --seed 42 --state run.json

The grade checks what actually changed: the right records were updated, the right people got exactly one message each, nothing outside the task was touched, and no money moved unless the task allowed it. See docs/run-with-claude-code.md for a full walkthrough.

Use the data on its own

npx @teamshift/fake-business --industry marketing-agency --seed 7 --format csv --out ./agency
npx @teamshift/fake-business --industry home-services --format sql-sqlite --out business.sql

import { generate } from "@teamshift/fake-business";

const biz = generate({ industry: "dental-clinic", seed: 7 });
const overdue = biz.invoices.filter((i) => i.status === "open" && i.dueOn < biz.meta.asOf);
console.log(biz.company.name, overdue.length, "overdue invoices");
console.log(biz.anomalies.map((a) => a.kind)); // ground truth for grading

FAQ

How do I test an AI agent without real customer data?

Shortened. The full README is on GitHub.