---
title: "About AgentTwin"
description: "AgentTwin helps developers test prepared application and AI-agent workflows with synthetic service data, supported faults, and explicit outcome checks."
canonical: "https://agenttwin.si/about/"
---

# About AgentTwin

AgentTwin helps developers test prepared application and AI-agent workflows against declared starting data, supported service operations, and expected outcomes. Its published Node.js CLI can discover repository dependencies and candidate code paths. Discovery helps prepare a workflow; it does not prove that a path is runnable or that a production environment has been reproduced.

## What AgentTwin does

A workflow may depend on external services whose behavior changes with timing and prior state. AgentTwin lets you describe a test world with initial data, supported operations, and assertions for the result you expect. Prepare a Node workflow, run it against the declared synthetic or bounded service behavior, and inspect its recorded outcome.

The declared world covers only the operations and conditions it contains. Local execution still has access to the host filesystem and network, so review the code, configuration, and fixtures before running it. Use synthetic data when practical.

## How it works

Install Node.js 22.12 or newer, then run `npx agenttwin init` in your repository to discover candidate services and prepare a workflow. You review the resulting files and decide what should run. The CLI supports prepared Node workflows using synthetic HTTP collections and bounded provider adapters. The included reference example demonstrates a clean run and a timeout-after-write fault that can reveal a duplicate write.

Run receipts record the supported interactions and assertion results observed by that run. The CLI can verify, reset, and replay supported examples. The local `npx agenttwin mcp` command exposes world, scenario, execution, and inspection tools to compatible coding agents over stdio; it is not a hosted workflow execution endpoint.

## Choosing a first test

Start with one small workflow whose result you can describe. For example, choose a write that should happen once even if a response is lost. Define the initial records, the action to run, and the final state you expect. Add a check for unwanted side effects as well as the required result. Keep the same fixtures when comparing a failing run with a proposed fix. This makes the comparison easier to review and gives you a focused regression to keep as the application changes.

## Boundaries and next steps

The browser demo is a scripted preview with synthetic data. Cloud execution is optional and requires E2B configuration. Neither discovery, a passing test, nor a receipt proves universal provider fidelity, security certification, or production equivalence. Read the [safety boundaries](/docs/safety/), [supported provider operations](/docs/provider-operations/), [quickstart](/docs/quickstart/), and [current status](/docs/status/) before choosing a workflow to test.

To discuss whether AgentTwin supports a workflow, [book a call](https://cal.com/sashanksingh).