Voice agents are a black box. You ship one and immediately wonder: which model is actually running, why is the first word slow, and what is this provider bill. I built VoiceGateway to answer those without changing your provider setup. One call, attach(session), and the profiler wraps your existing STT, LLM, and TTS plugins. It never sits in the audio path. It records latency (TTFT, first audio, p50/p95), the exact model behind each call, and cost per modality. Storage is local SQLite. Your keys, your data. Latest version v0.21.1 adds three things: per-agent live compute and memory in the dashboard, a cached latency probe (one real measured call, then cached), and a probe on the Overview card that places one real call to an agent so you can verify end to end before a deploy. Live demo, no login: https://voicegateway.dev/demo What would you find most useful to see profiled first? Happy to answer anything.

VoiceGateway

The open source profiler for voice agents

The open source profiler for voice agents