> ## Documentation Index
> Fetch the complete documentation index at: https://docs.argide.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Overview

> A headless runtime that lets an Argide agent drive your web app

The Agent SDK (`@ourguide-ai/agent`) loads a headless agent runtime in your end user's browser tab. An Argide agent can then drive that tab — clicking, filling forms, navigating — to complete a goal your own chatbot delegates to it.

Where the [chat widget](/setup/embed-widget) gives you a complete, Argide-owned chat UI, the Agent SDK ships no UI at all. Your product keeps its own assistant; Argide takes over only when a conversation turns into something that must be *done* in the app rather than answered.

## How it works

```mermaid theme={null}
sequenceDiagram
    participant Page as Your page (SDK)
    participant Server as Your server
    participant Argide as Argide backend
    participant Agent as Argide agent

    Page->>Argide: init() — mint session
    Argide-->>Page: { sessionId, pairingCode }
    Page->>Server: forward the pair
    Server->>Argide: claim (Authorization: Bearer ak_live_…)
    Argide-->>Server: session claimed
    Page->>Argide: do({ goal: "cancel my subscription" })
    Argide->>Agent: run task
    Agent->>Page: clicks, fills forms, navigates
    Page->>Argide: check({ taskId })
    Argide-->>Page: { status, summary }
```

1. **The SDK mints a session.** `init()` runs in the tab and receives `{ sessionId, pairingCode }`.
2. **Your server claims it.** The browser forwards the pair to your backend, which exchanges it for a claim using your `ak_live_...` API key. Until then, the session can't run tasks.
3. **Your chatbot delegates a goal.** `do({ goal })` hands the agent a job and resolves with a `taskId` as soon as the task exists.
4. **The agent drives the tab.** Clicking, filling forms, navigating — in the user's own session, right in front of them.
5. **You read the outcome.** `check({ taskId })` returns the authoritative `{ status, summary }`.

## Security model

* **The browser never holds your API key.** The `ak_live_...` key lives on your server; the tab only ever sees a single-use pairing code.
* **Unclaimed sessions can't act.** Tasks are refused until your server has claimed the session.
* **A pairing code is exchanged once.** Re-claiming an already-claimed session returns `409 ALREADY_CLAIMED` — reserved as the signal that a stolen code is being replayed.

[Claiming Sessions](/agent-sdk/claiming-sessions) covers this in depth.

## Integration requirements

* **Package access.** The SDK is a restricted package on GitHub Packages, so you need an access grant and an authenticated token before you can install it — see [Installation](/agent-sdk/installation).
* **Client-side tool execution.** Your stack must be able to execute tools in the browser.

## Get started

<CardGroup cols={2}>
  <Card title="Installation" icon="key" href="/agent-sdk/installation">
    Get package access and configure your token
  </Card>

  <Card title="Quickstart" icon="rocket" href="/agent-sdk/quickstart">
    Initialize, claim, and run your first task
  </Card>
</CardGroup>