---
title: "Welcome to AgentTwin"
description: "The handbook from repository discovery to a reproduced failure and a verified application fix."
canonical: "https://agenttwin.si/docs/"
---

# Welcome to AgentTwin

The handbook from repository discovery to a reproduced failure and a verified application fix.


AgentTwin discovers project dependencies and workflow candidates, then tests reviewed Node workflows through the CLI or an optional cloud runner. Inspect assertions, service state, logs and failures before changing production.

## Start with your project

Run this inside your repository with Node.js 22.12 or newer:

```sh
npx agenttwin init
```

Review the generated discovery. Init can offer an installed OpenCode session; other coding agents can use `.agenttwin/AGENT-TASK.md`. Analysis identifies candidates, required dependencies and missing inputs. It does not prove a workflow works.

Connect your individual AgentTwin account in [the dashboard](/dashboard/). Run `npx agenttwin login`, complete the browser connection, then run `npx agenttwin sync`. Automation can supply `AGENTTWIN_TOKEN` to send reviewed discovery metadata. You can also import `.agenttwin/inventory.json` manually.

<CardGroup cols={2}>
  <Card title="Workflow guide" href="/workflow">Discover, prepare, execute and inspect your actual code.</Card>
  <Card title="Dashboard guide" href="/dashboard">Environment, Workflows, Runs, Settings and the account menu.</Card>
  <Card title="Current capabilities" href="/status">Supported contracts and remaining gaps.</Card>
  <Card title="Hosted execution" href="/safety">Understand starting state, credentials and sandbox boundaries.</Card>
</CardGroup>

## Coverage is explicit

The runtime supports custom HTTP collections, a bounded Slack API subset, PostgreSQL and Redis. Catalog selection declares a dependency; it does not implement a vendor adapter or create a real provider account. Unsupported behavior must be reviewed before execution.

The product covers many workflows, including support, notifications, sales, payments and agent side effects. Refunds are an example, not the product boundary.

## Read the handbook from start to finish

| Chapter | What you will learn |
| --- | --- |
| [First verified fix](/docs/quickstart/) | Execute an example failure and prove a source-only repair |
| [Discover and test](/docs/workflow/) | Move from repository signals to reviewed execution |
| [Prepare application code](/docs/preparation/) | Build an entrypoint, configure twins and review the bundle |
| [Reusable worlds](/docs/worlds/) | Provision, run, reset, extend and clean up |
| [Scenarios and fixtures](/docs/scenarios/) | Link synthetic identities and preserve immutable baselines |
| [Fidelity](/docs/fidelity/) | Understand operation coverage and provider limitations |
| [Safety](/docs/safety/) | Choose local or cloud boundaries and review credentials |
| [Evidence and retesting](/docs/evidence/) | Diagnose failures and compare unchanged checks |
| [CLI and MCP](/docs/cli/) | Find commands, artifacts and coding-agent tools |
| [Troubleshooting](/docs/troubleshooting/) | Resolve setup, account and execution failures |

For a first-time user, start with the local quickstart before connecting a real project. For an existing workflow, begin with preparation. For a failed run, begin with evidence. The handbook describes current implemented contracts; planned features remain labeled in the roadmap.