Python workspace for Smelter, with:

• lib/ — smelter-sdk, the published client library.
• example/ — three runnable demos that exercise Smelter's side channel.

The SDK exposes Smelter's side channel: a Unix-socket stream of decoded RGBA video frames and PCM audio batches, so Python code can run ML inference (YOLO, Whisper, …) against live media and feed results back via Smelter's HTTP API.

cd sdks/python
uv sync

Minimal side-channel subscribers — enumerate every published channel and print one line per video frame and per audio batch. Use these as a starting point when wiring the SDK into your own code.

Requires a Smelter server already running with side_channel enabled on at least one input. Point the env var at the directory Smelter writes its sockets into:

SMELTER_SIDE_CHANNEL_SOCKET_DIR=/path/to/sockets uv run example-simple-sync
SMELTER_SIDE_CHANNEL_SOCKET_DIR=/path/to/sockets uv run example-simple-async

Full demo: boots a Smelter server, runs YOLO object detection and Whisper speech-to-text on the input, and renders detection boxes and subtitles back over the stream.

uv run example-yolo-whisper <optional_example.mp4>

By default it detects people. Override the YOLO class filter via env var:

DETECT_CLASSES=car,truck uv run example-yolo-whisper ./cars.mp4

If you don't pass an MP4, the example registers a WHIP input. Stream your webcam to it from:

https://smelter-labs.github.io/tools/#whip-streamer?url=http://127.0.0.1:9000/whip/input&token=example

The example exposes a WHEP endpoint with detection boxes + subtitles:

https://smelter-labs.github.io/tools/#whep-player?url=http://127.0.0.1:9000/whep/output&token=example

It also writes a fragmented MP4 to /tmp/smelter_output.mp4.