
Sixty containers on one server
One bare-metal box runs dozens to hundreds of Hoody containers. KSM and BTRFS dedup make the marginal cost near zero.
You're pair-debugging at midnight and the other side's hotspot won't survive a Zoom. You both have a shell. ffmpeg captures the mic, curl PUTs it to a pipe path, the other terminal curls and plays the bytes. No SDK, no signaling, no meeting.
$ ffmpeg -f alsa -i default \
-c:a libopus -f ogg - \
| curl -T - https://pipe.containers.hoody.com/api/v1/pipe/voice$ curl https://pipe.containers.hoody.com/api/v1/pipe/voice
| mpv --no-video -no signaling, no SDK, no Zoom
ffmpeg encodes your microphone to opus and writes it to stdout. curl PUTs that stdin to /api/v1/pipe/voice. Whoever GETs the same path receives the bytes as the sender produces them. The pipe holds the receiver up to five minutes for the sender to connect, forwards Content-Type, and stores nothing.
ffmpeg -f alsa -i default -c:a libopus -f ogg - reads your microphone, encodes to opus, writes to stdout.
curl -T - PUTs stdin to /api/v1/pipe/voice. The pipe waits up to five minutes for a receiver on the same path.
Their terminal runs curl https://.../api/v1/pipe/voice. The pipe matches sender to receiver and starts streaming.
Their curl pipes into aplay or mpv. Audio plays as it arrives, byte for byte. ctrl-C ends the call.
The pipe forwards Content-Type, supports up to five-minute hold-time before a counterpart connects, and uses HTTPS — nothing more exotic than the protocol your shell already speaks.
There is no echo cancellation, no jitter buffer, no fancy DSP. It is the kind of voice channel an SRE pair-debugs over: stripped down, low overhead, dependable. Every layer a meeting platform builds is a layer the pipe pair simply doesn't have.
The same mic-over-HTTP mechanism reads three different ways depending on who is on the other end of the pipe.
You're three commands deep into a prod incident. Spinning up Zoom would take longer than the fix. ffmpeg | curl on your side, curl | mpv on theirs — you can talk while you tail logs.
Phone-tethering chokes on a video call. A 32 kbps opus stream over HTTP doesn't. The other end opens a URL and listens — they don't even need a mic to participate.
Nothing is stored on a server. There is no third-party app on either machine. The pipe is purely streaming — bytes pass through and are gone the moment ctrl-C lands.
Audio is just bytes. Bytes are just a pipe.
The lineup of voice tooling every pair of engineers accumulates. Each one assumes a meeting, an account, or a custom signaling server. The pipe URL assumes none of those.
The next time someone says 'can we hop on a quick call', open a pipe instead.