Real-time voice agents get a native Rust runtime in Flowcat: it carries a phone or WebRTC call through a composable media pipeline on your own infrastructure, so latency-sensitive audio never queues behind a managed platform. Audio flows from transport into voice activity detection and turn-taking, then through speech-to-text, an LLM, and text-to-speech, or a single speech-to-speech model, and back out to the caller, with in-process SIP and RTP for telephony. Pair it with local Whisper transcription, Kokoro, Piper, or XTTS for speech, and Ollama for reasoning, and a call's audio and transcript never leave your network; cloud providers plug in by name when you prefer them. Describe one agent in a YAML or JSON node-and-edge graph and the single flowcat-server binary serves it over HTTP, opening a browser page with microphone and live transcript, or bridging a Plivo number into the same agent. The runtime stays flat as call volume grows: benchmarked against a multiprocess Python baseline, one process holds p99 round-trip overhead at or below 0.61 milliseconds from 10 to 2,000 concurrent calls, at roughly 19.6 KB of RAM per idle session. Deploy it inside your VPC or an air-gapped environment where call recordings are sensitive. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.