Voice & Browser Agent Examples

import { VoiceAgent, openai } from "@agentium/core";

const agent = new VoiceAgent({
  name: "voice-greeter",
  model: openai("gpt-4o-realtime"),
  instructions: "You are a friendly voice assistant. Speak naturally and keep answers concise.",
  voice: "alloy",
  inputFormat: "pcm16",
  outputFormat: "pcm16",
  sampleRate: 24000,
});

const session = await agent.connect();

session.on("speech", (audio) => {
  // Stream PCM16 audio to speaker
  speaker.write(audio);
});

session.on("transcript", (text, role) => {
  console.log(`[${role}] ${text}`);
});

// Pipe microphone input to the agent
microphone.on("data", (chunk) => session.send(chunk));

// Graceful shutdown
process.on("SIGINT", async () => {
  await session.close();
  process.exit(0);
});

import { VoiceAgent, defineTool, openai } from "@agentium/core";
import { z } from "zod";

const lookupOrder = defineTool({
  name: "lookup_order",
  description: "Look up an order by ID and return its status",
  parameters: z.object({
    orderId: z.string().describe("The order ID, e.g. ORD-1234"),
  }),
  execute: async ({ orderId }) => {
    const orders: Record<string, string> = {
      "ORD-1234": "Shipped — arrives Friday",
      "ORD-5678": "Processing — estimated Monday",
    };
    return orders[orderId] ?? `No order found for ${orderId}`;
  },
});

const agent = new VoiceAgent({
  name: "voice-support",
  model: openai("gpt-4o-realtime"),
  instructions:
    "You are a customer support agent. Look up orders when asked and read back the status clearly.",
  voice: "shimmer",
  tools: [lookupOrder],
  toolCallBehavior: "speakBeforeAndAfter",
});

const session = await agent.connect();

session.on("speech", (audio) => speaker.write(audio));
session.on("toolCall", (name, args) => {
  console.log(`[tool] ${name}(${JSON.stringify(args)})`);
});

microphone.on("data", (chunk) => session.send(chunk));

import { VoiceAgent, InMemoryStorage, openai } from "@agentium/core";

const storage = new InMemoryStorage();

const agent = new VoiceAgent({
  name: "voice-personal",
  model: openai("gpt-4o-realtime"),
  instructions:
    "You are a personal assistant. Remember the user's name, preferences, and past requests across sessions.",
  voice: "echo",
  memory: {
    storage,
    memoryKeys: ["user_name", "preferences", "past_requests"],
    autoSummarize: true,
  },
});

const userId = "user-42";

// Session 1: user introduces themselves
const s1 = await agent.connect({ sessionId: `${userId}-session-1` });
s1.on("speech", (audio) => speaker.write(audio));
microphone.on("data", (chunk) => s1.send(chunk));
// User says: "Hi, I'm Sarah. I prefer metric units."
// Agent remembers name and preference.

// Later — Session 2: agent already knows the user
const s2 = await agent.connect({ sessionId: `${userId}-session-2` });
s2.on("speech", (audio) => speaker.write(audio));
microphone.on("data", (chunk) => s2.send(chunk));
// User says: "What's the weather?"
// Agent: "Hi Sarah! It's 22°C and sunny in your area." (uses metric without asking)

import { VoiceAgent, CostTracker, openai } from "@agentium/core";

const costTracker = new CostTracker({
  pricing: {
    "gpt-4o-realtime": {
      promptPer1k: 0.005,
      completionPer1k: 0.02,
      audioInputPer1k: 0.06,
      audioOutputPer1k: 0.24,
    },
  },
});

const agent = new VoiceAgent({
  name: "voice-tracked",
  model: openai("gpt-4o-realtime"),
  instructions: "You are a concise voice assistant.",
  voice: "alloy",
  costTracker,
});

const session = await agent.connect();
session.on("speech", (audio) => speaker.write(audio));
microphone.on("data", (chunk) => session.send(chunk));

session.on("turnComplete", () => {
  const summary = costTracker.getSummary();
  console.log("--- Cost Report ---");
  console.log("Audio input tokens:", summary.totalTokens.audioInputTokens);
  console.log("Audio output tokens:", summary.totalTokens.audioOutputTokens);
  console.log("Text tokens:", summary.totalTokens.promptTokens + summary.totalTokens.completionTokens);
  console.log("Total cost:", `$${summary.totalCost.toFixed(4)}`);
});

// After a 2-minute conversation:
// Audio input tokens:  14400
// Audio output tokens: 9600
// Text tokens:         320
// Total cost:          $3.18

import { VoiceAgent, openai, defineTool } from "@agentium/core";
import { VoiceGateway } from "@agentium/core";
import { z } from "zod";

const checkBalance = defineTool({
  name: "check_balance",
  description: "Check the caller's account balance",
  parameters: z.object({
    accountId: z.string().describe("Caller's account ID"),
  }),
  execute: async ({ accountId }) => {
    return `Account ${accountId}: balance $142.50`;
  },
});

const agent = new VoiceAgent({
  name: "phone-support",
  model: openai("gpt-4o-realtime"),
  instructions: "You are a phone banking assistant. Verify the account before giving balance info.",
  voice: "shimmer",
  tools: [checkBalance],
});

const gateway = new VoiceGateway({
  agent,
  port: 8080,
  path: "/voice",
  audioFormat: "pcm16",
  sampleRate: 24000,
  maxConcurrentSessions: 50,
  onConnect: (ws, req) => {
    console.log(`New voice connection from ${req.socket.remoteAddress}`);
  },
  onDisconnect: (sessionId, duration) => {
    console.log(`Session ${sessionId} ended after ${duration}ms`);
  },
});

await gateway.start();
console.log("Voice gateway listening on ws://localhost:8080/voice");

// Client connects via WebSocket and streams raw audio frames.
// The gateway handles session management, VAD, and audio routing.

import { Agent, openai } from "@agentium/core";
import { BrowserAgent } from "@agentium/browser";

const agent = new BrowserAgent({
  name: "web-scraper",
  model: openai("gpt-4o"),
  instructions: "Navigate to websites and extract the requested information accurately.",
  headless: true,
  viewport: { width: 1280, height: 720 },
});

const result = await agent.run(
  "Go to https://news.ycombinator.com and extract the top 5 story titles with their point counts."
);

console.log(result.text);
// → "1. Show HN: Open-source LLM framework (342 points)
//    2. The future of quantum computing (289 points)
//    3. ..."

await agent.close();

import { Agent, openai } from "@agentium/core";
import { BrowserAgent } from "@agentium/browser";

const agent = new BrowserAgent({
  name: "form-filler",
  model: openai("gpt-4o"),
  instructions: `
    Fill out web forms with the provided data.
    Always verify the form was submitted successfully before reporting back.
  `,
  headless: false,
  viewport: { width: 1280, height: 800 },
  defaultTimeout: 15000,
});

const formData = {
  firstName: "Alice",
  lastName: "Johnson",
  email: "alice@example.com",
  company: "Acme Corp",
  role: "Engineering Manager",
  message: "Interested in the enterprise plan. Please contact me.",
};

const result = await agent.run(
  `Go to https://example.com/contact and fill out the contact form with this data: ${JSON.stringify(formData)}. Submit the form and confirm it went through.`
);

console.log(result.text);
// → "Successfully submitted the contact form. Confirmation message displayed: 'Thank you, Alice! We'll be in touch within 24 hours.'"

await agent.close();

import { Agent, openai } from "@agentium/core";
import { BrowserAgent } from "@agentium/browser";
import { writeFile } from "node:fs/promises";

const agent = new BrowserAgent({
  name: "visual-analyzer",
  model: openai("gpt-4o"),
  instructions: "Navigate to pages, take screenshots, and analyze visual content in detail.",
  headless: true,
  viewport: { width: 1440, height: 900 },
  screenshotFormat: "png",
});

const result = await agent.run(
  "Go to https://example.com/dashboard, take a screenshot, and describe the layout and key metrics visible."
);

console.log("Analysis:", result.text);

if (result.screenshots?.length) {
  for (const [i, screenshot] of result.screenshots.entries()) {
    await writeFile(`screenshot-${i}.png`, screenshot.buffer);
    console.log(`Saved screenshot-${i}.png (${screenshot.width}x${screenshot.height})`);
  }
}

await agent.close();

import { Agent, CostTracker, openai } from "@agentium/core";
import { BrowserAgent } from "@agentium/browser";

const costTracker = new CostTracker({
  pricing: {
    "gpt-4o": {
      promptPer1k: 0.0025,
      completionPer1k: 0.01,
      imagePer1k: 0.003825,
    },
  },
});

const agent = new BrowserAgent({
  name: "tracked-browser",
  model: openai("gpt-4o"),
  instructions: "Browse and extract information from websites.",
  headless: true,
  costTracker,
});

const result = await agent.run(
  "Go to https://example.com/pricing, take a screenshot, and list all pricing tiers with their features."
);

console.log(result.text);

const summary = costTracker.getSummary();
console.log("\n--- Browser Agent Cost Report ---");
console.log("LLM calls:", summary.totalCalls);
console.log("Prompt tokens:", summary.totalTokens.promptTokens);
console.log("Completion tokens:", summary.totalTokens.completionTokens);
console.log("Image tokens (screenshots):", summary.totalTokens.imageTokens ?? 0);
console.log("Total cost:", `$${summary.totalCost.toFixed(4)}`);
// → Total cost: $0.0847 (3 screenshots analyzed)

await agent.close();

import { Agent, openai } from "@agentium/core";
import { BrowserAgent } from "@agentium/browser";
import { writeFile } from "node:fs/promises";

const agent = new BrowserAgent({
  name: "workflow-agent",
  model: openai("gpt-4o"),
  instructions: `
    You are an automation agent. Follow multi-step workflows precisely:
    1. Complete each step before moving to the next
    2. Verify success at each stage
    3. If a step fails, retry once before reporting the error
    4. Extract all requested data in structured format
  `,
  headless: true,
  viewport: { width: 1280, height: 800 },
  defaultTimeout: 30000,
  downloadPath: "./downloads",
});

const result = await agent.run(`
  Complete this workflow on https://app.example.com:

  Step 1 — Login:
    - Username: demo@example.com
    - Password: ${process.env.DEMO_PASSWORD}

  Step 2 — Navigate to Reports:
    - Click "Reports" in the sidebar
    - Select "Monthly Revenue" report

  Step 3 — Extract Data:
    - Read the revenue figures for the last 3 months
    - Note any month-over-month changes

  Step 4 — Download:
    - Click the "Export CSV" button
    - Wait for the download to complete

  Return the extracted data as JSON.
`);

console.log("Extracted data:", result.text);
// → { "months": [
//      { "month": "Jan 2025", "revenue": "$142,500", "change": "+12%" },
//      { "month": "Feb 2025", "revenue": "$158,200", "change": "+11%" },
//      { "month": "Mar 2025", "revenue": "$167,900", "change": "+6%" }
//    ] }

if (result.downloads?.length) {
  console.log("Downloaded files:", result.downloads.map((d) => d.filename));
}

await agent.close();