Appearance

Site Agent Copilot ​

Turn any web application into an AI-navigable workspace. Users describe what they want; the Agent does it.

The Problem ​

Your users get lost in complex UIs. They open support tickets asking "how do I change my billing plan?" when the answer is three clicks away. Traditional chatbots can answer questions but can't do anything.

The Solution ​

Site Agent Copilot is an AI assistant that lives inside your web app and can:

  • Answer questions from a knowledge base (pricing, features, how-to)
  • Operate the page on behalf of users — navigate, click, type, select
  • Guide visually — highlights elements before clicking, reports progress in chat

No API integration required. No backend changes. The Agent operates through the DOM with the same permissions as the logged-in user.

How It Looks ​

User: "Help me create a new project"

Agent: Navigating to Projects page...
       ✓ Navigated to /projects
       Clicking "New Project"...
       ✓ Clicked
       Filling in project name...
       ✓ Done

Agent: I've created a blank project for you. 
       Want me to configure the settings too?

The user watches their page being operated in real-time, with elements briefly highlighted before each action.

Architecture ​

┌──────────────────────────────────────────────┐
│  Browser                                      │
│  Chat Widget ←─ SSE ─→ Frontend Agent (DOM)   │
└───────────────────────┬──────────────────────┘
                        │ Session
┌───────────────────────┴──────────────────────┐
│  SandBase Cloud                               │
│  Embed API → Cloud Agent (Hermes + Knowledge) │
└──────────────────────────────────────────────┘
  • Cloud Agent thinks and plans (runs in a sandbox)
  • Frontend Agent executes DOM actions (runs in browser)
  • Session connects them (SSE downstream, POST upstream)

Use Cases ​

SaaS Onboarding ​

Replace static product tours with a conversational guide:

  • "Help me set up my workspace" → Agent navigates through setup wizard
  • Adapts to what the user already configured vs. what's missing
  • Works in any language the LLM supports

Enterprise Admin Panels ​

Complex admin UIs with dozens of nested menus:

  • "Set the production log level to debug" → Agent finds the right page, changes the dropdown, saves
  • Faster than teaching users where every setting lives

ERP / CRM Form Automation ​

Multi-step data entry:

  • "Create an invoice for Acme Corp, $5,000, due in 30 days" → Agent fills the entire form
  • Handles multi-page workflows that span several routes

Accessibility ​

Users who can't easily navigate with a mouse:

  • Voice input (via STT) → Agent operates the page
  • Screen-reader-friendly: the Agent describes what it's doing

Internal Tools Support ​

IT help desk for internal tools:

  • Employees ask the copilot instead of filing tickets
  • Agent either answers directly or performs the action for them

Getting Started ​

Full guide: Site Agent integration

Quick version:

  1. Create a SandBase Agent with frontend_action tool + knowledge skills
  2. Create an Embed Config → get a publishable key
  3. Add the widget to your frontend (one React component or one <script> tag)
  4. Write a knowledge base describing your pages and operations

Time to first working demo: ~30 minutes.

Cost ​

  • Cloud Agent reasoning: standard SandBase model pricing (e.g., Claude Sonnet 4 at $3/$15 per 1M tokens)
  • Frontend Agent (DOMExecutor): free — runs entirely client-side
  • Frontend Agent (PageAgentExecutor): ~$0.002 per action step (one LLM call for element location)

A typical 5-step page operation costs $0.01–$0.05 total.

Limitations ​

  • Only operates same-origin pages (can't navigate to external sites)
  • React-controlled inputs may need custom event dispatching
  • Dynamic content that loads after navigation needs a brief wait
  • Not a replacement for proper UX — use it to supplement, not substitute