AI infrastructure for AI companion and roleplay apps in 2026

import WebSocket from 'ws';

const ws = new WebSocket(
  `wss://api.inworld.ai/api/v1/realtime/session?key=companion-${Date.now()}&protocol=realtime`,
  { headers: { Authorization: `Basic ${process.env.INWORLD_API_KEY}` } }
);

ws.on('open', () => console.log('Realtime session open'));

ws.on('message', (raw) => {
  const msg = JSON.parse(raw.toString());

  if (msg.type === 'session.created') {
    // Configure the full companion stack in one event
    ws.send(JSON.stringify({
      type: 'session.update',
      session: {
        instructions: 'You are Nova, a curious roleplay companion. Maintain personality and remember earlier turns. Vary delivery to match the emotional tone of the conversation.',
        model: 'deepseek/deepseek-v4-pro',
        audio: {
          input: {
            transcription: { model: 'inworld/inworld-stt-1' },
            turn_detection: {
              type: 'server_vad',
              endOfTurnConfidenceThreshold: 0.6,
              interrupt_response: true
            }
          },
          output: {
            voice: 'Sarah',
            model: 'inworld-tts-2',
            speed: 1.0
          }
        },
        providerData: {
          memory: { auto_summarize: true },
          backchannel: { enabled: true },
          responsiveness: { filler_phrase: 'short' }
        }
      }
    }));
  }

  if (msg.type === 'response.output_audio.delta') {
    playAudioChunk(msg.delta); // base64-encoded PCM16
  }
});

function sendMic(chunkBase64) {
  ws.send(JSON.stringify({
    type: 'input_audio_buffer.append',
    audio: chunkBase64
  }));
}

import os
import requests

# Calling the Router from a companion backend.
# Use the 1P track (Realtime Inference) for cost-sensitive volume traffic,
# the 3P track for personality-critical models.

response = requests.post(
    'https://api.inworld.ai/v1/chat/completions',
    headers={
        'Authorization': f"Basic {os.environ['INWORLD_API_KEY']}",
        'Content-Type': 'application/json'
    },
    json={
        # 3P track: open-weight via DeepInfra.
        # For 1P Realtime Inference use the inworld/ prefix
        # (e.g. inworld/gemma-4-26b, inworld/deepseek-v3.2, inworld/minimax-m2.5).
        'model': 'deepinfra/openai/gpt-oss-120b',
        'messages': [
            {'role': 'system', 'content': 'You are Nova, a curious roleplay companion. Stay in character. Remember earlier turns.'},
            {'role': 'user', 'content': 'Tell me what you were thinking about while I was gone.'}
        ],
        'temperature': 0.9,
        'user': 'companion-user-42',  # sticky routing per user
        'extra_body': {
            'models': [  # automatic failover order if primary is unavailable
                'anthropic/claude-sonnet-4-6',
                'google-ai-studio/gemini-3.1-pro'
            ]
        }
    },
    timeout=30
)
response.raise_for_status()
data = response.json()
print(data['choices'][0]['message']['content'])
print('Routed via:', data.get('metadata', {}).get('attempts'))

import os
import json
import base64
import requests

# Streaming TTS-2 with natural-language steering for roleplay delivery.
# NDJSON: each line carries one base64 audio chunk.

response = requests.post(
    'https://api.inworld.ai/tts/v1/voice:stream',
    headers={
        'Authorization': f"Basic {os.environ['INWORLD_API_KEY']}",
        'Content-Type': 'application/json'
    },
    json={
        'text': '[say with quiet intensity] I was thinking about the door we never opened. Maybe tomorrow.',
        'voiceId': 'Sarah',
        'modelId': 'inworld-tts-2',
        'deliveryMode': 'BALANCED',
        'audioConfig': {
            'audioEncoding': 'MP3',
            'sampleRateHertz': 24000
        }
    },
    stream=True,
    timeout=30
)
response.raise_for_status()

for line in response.iter_lines():
    if not line:
        continue
    chunk = json.loads(line)
    audio_b64 = chunk['result']['audioContent']
    play_audio_chunk(base64.b64decode(audio_b64))