Talk naturally with specialized AI agents in a calm, minimal video-call style room.

Arlo lets you have a real spoken conversation with an AI agent — no typing, no prompts. Pick an agent, click Start, and just talk. The agent listens, thinks, and responds in a natural human voice. It knows when you're done speaking and won't cut you off.

Agents:

• CMSC420 TA — Advanced Data Structures (trees, heaps, graphs, spatial indexing, complexity)
• CMSC417 TA — Computer Networks (TCP, DNS, routing, congestion control, protocols)
• Telehealth Intake — Calm, structured pre-visit intake assistant
• Interview Coach — Behavioral and technical interview practice

Features:

• Voice conversation with natural turn-taking — no button to press when done
• Human-sounding ElevenLabs voices via Vapi
• Visual avatar that animates while listening, thinking, and speaking
• Auto-captured session notes + full transcript
• Paste document URLs (Google Docs, lecture notes, course pages) as context so the agent knows your specific course material

npm install

Vapi (handles voice pipeline — STT, LLM, TTS, turn-taking)

1. Sign up at vapi.ai — free $10 credit on signup
2. Go to Account → copy your Public Key

Anthropic (Claude powers the AI responses)

1. Get a key at console.anthropic.com

cp .env.example .env.local

Fill in .env.local:

ANTHROPIC_API_KEY=sk-ant-...
NEXT_PUBLIC_VAPI_PUBLIC_KEY=pk_...

npm run dev

Open the URL shown in your terminal (usually http://localhost:3000).

Note: Use Chrome or Edge — they have the best Web Speech API support.

Vapi chains together three services into a seamless real-time call:

1. Deepgram Nova-2 — streaming speech-to-text with smart VAD (knows when you're done)
2. Claude Haiku — fast, concise AI responses tuned for spoken conversation
3. ElevenLabs — natural-sounding TTS voices, one per agent

Each agent has a distinct human face avatar that reacts in real time:

• Idle — gentle breathing animation
• Listening — ripple rings expand outward
• Thinking — face dims, dots bounce below
• Speaking — mouth opens and closes, waveform pulses

On the agent selection page, paste up to 3 URLs before starting a session. Arlo fetches the text and injects it into the agent's context — so it can answer questions about your specific lecture notes or class materials.

Supported sources:

• Google Docs (set to "Anyone with link can view")
• Google Slides
• Any public HTML page (course websites, lecture notes, wikis)

The session room has a right panel with two tabs:

• Notes — auto-captures a key point from each agent response, plus free-form typing area
• Transcript — full back-and-forth conversation log

app/
  page.tsx                 # Landing page
  select/page.tsx          # Agent + document selection
  session/page.tsx         # Live call room
  api/chat/route.ts        # Claude API endpoint (fallback)
  api/fetch-doc/route.ts   # Document URL fetcher

components/
  Avatar.tsx               # Animated human face avatar (4 styles, 4 states)
  AgentCard.tsx            # Selectable agent card
  DocumentInput.tsx        # URL input for study materials
  Notepad.tsx              # Auto-capture + freeform notes panel
  Transcript.tsx           # Conversation log
  StatusPill.tsx           # Idle / Listening / Thinking / Speaking indicator
  ControlBar.tsx           # Mic / End call controls

lib/
  personas.ts              # Agent definitions, system prompts, voice IDs
  claude.ts                # /api/chat client helper
  speech.ts                # Browser Web Speech API utilities
  turnTaking.ts            # Pause detection logic (fallback)

types/
  speech.d.ts              # Web Speech API type declarations