Introduction

autopilot-client is a TypeScript library for connecting to LLMOR autopilot servers over WebSockets. It is the browser-side counterpart that lets a page or app participate in a conversation that the server orchestrates: streaming completions, calling client-side functions, syncing UI context, and rendering a chat surface for the user.

What you get

The package exposes three layers, each building on the one below it.

1. AutopilotClient — headless and framework-agnostic

A single class that:

• Fetches conversation metadata over HTTP and opens a Socket.IO connection to the relay.
• Lets you register functions the autopilot can call (with JSON-Schema parameters and async handlers).
• Lets you sync arbitrary context state to the server (debounced JSON Patch diffs).
• Emits typed events for status changes, streaming chunks, function calls, and conversation updates.
• Has no dependency on Vue or any UI framework — works in Node, in a vanilla page, or inside any framework.

→ Client overview

2. Vue + Pinia store — reactive bridge

useAutopilotStore wires every relevant event from AutopilotClient into a Pinia store, so your Vue components get reactive connectionStatus, messages, streamingMessage, context, and a curated visibleMessages computed.

A companion composable useAutopilotConversation exposes filtered message lists (user, assistant, system) for building custom chat UIs.

→ Vue overview

3. AutopilotBar — drop-in chat UI

Call enableAutopilotBar(client) once and you get a floating toggle button plus an overlay chat surface, mounted inside a Shadow DOM so its Tailwind styles do not leak into your host page.

→ Bar overview

When to use which

You want to…	Use
Build a custom chat UI in React, Svelte, plain JS, or Node	Headless AutopilotClient
Build a custom chat UI in Vue 3	Headless client + Pinia store
Add an autopilot to an existing page without designing UI	AutopilotBar
Embed an autopilot inside an existing Vue 3 + Pinia app, with no custom UI	AutopilotBar inside a Vue app

What this library is not

• It is not an LLM client. All inference happens on the LLMOR server; this package only relays.
• It is not a UI framework. The bundled AutopilotBar is a single drop-in widget — it is not a library of design-system components.
• It does not implement authentication. You obtain a conversation URL (with token) from your backend and hand it to connect().

Next steps

• Install the package
• Quick start — three minimal snippets, one per layer
• Architecture — how the wire protocol fits together