Twins: the tools your agents work in, rebuilt for testing.
A twin is a stateful copy of a real tool, built for testing agents. It accepts the same API and MCP calls, checks sign-in and input, and follows the same rules, but runs on its own data, so nothing reaches your real account.
The real tool’s behaviour, on data that is yours to break.
Three things make a twin different from a mock and from a vendor sandbox.
The same interface
The same API and MCP tools as the real product, with the same sign-in. Your agent points at the twin’s address and runs unchanged.
The same rules
Required fields, permissions, rate limits, malformed input, the vendor’s real error types. A mistake fails in the twin the way it would fail in production, so you find it here.
Its own data, reset in one call
Private to your team and seeded with a company that already has history. Every run starts from the same state, and one call puts it back there.
Ten twins, across the tools your agents already use.
Each one covers the parts of the tool an agent actually touches, and each has its own page.
Don’t see the tool your agent uses? We build twins of internal tools and vendor APIs on request.
Request a twinEvery twin, out of the box.
Point your agent’s MCP config at the twin instead of the live tool. Nothing else in your agent changes.
{
"mcpServers": {
"salesforce": {
"type": "http",
"url": "<twin MCP endpoint>",
"headers": { "Authorization": "Bearer <token>" }
}
}
}REST works on the same data. Both are reset with one call.
Multiple Twins. One Company
The twins share their people, customers and history. A deal in Salesforce has the email threads, the meetings, the Slack channel and the Gong calls behind it, so an agent that updates the stage can be checked against what actually happened.
That is what makes a task that touches three tools run the way it would at work, and what lets the report grade the end state instead of the agent’s own summary.
- Opportunity: Acme renewal$48,000 · Negotiation · owner Mira Patelsalesforce
- Thread: “Renewal terms for next year”6 messages with the buyergmail
- Meeting: Acme pricing callThursday, 30 min, 3 attendeescalendar
- Call: Acme pricing callTranscript, next steps, two risks flaggedgong
- #deal-acmeEscalation thread, 14 messagesslack
Twins vs mocks vs sandboxes
Three ways to give an agent something to work in. Only one behaves like the real tool and resets like a test.
| Area | MocksStubs you write by hand | Vendor sandboxesThe vendor’s own dev org | Drizz twinsA stateful copy of the tool |
|---|---|---|---|
| Keeps state across steps | No | Yes | Yes |
| Real errors and rules | Only what you code | Yes | Yes |
| Connected across tools | No | No | One company, every twin |
| Reset between runs | Nothing to reset | Slow or manual | One call |
| Licences and limits | None | Paid org per tool, shared limits | None per tool, private |
| Graded report per run | No | No | Yes |
The long version: Mocks vs vendor sandboxes vs twins for testing AI agents →