auvux 0.1.0a5

Annotate faster. See deeper. Sound smarter.

auvux

Annotate faster. See deeper. Sound smarter.

---

Push inference runs from a Python notebook to a browser viewer over WebRTC.

Bytes flow peer-to-peer over a DataChannel; identical bytes across runs and viewers are deduplicated via per-viewer SHA-256 content addressing. Auth is GitHub OAuth (user:email scope) verified by an HMAC-signed JWT — your tokens never leave your machine.

Install

pip install --pre auvux

(--pre while we're on the 0.1.0a1 alpha release.)

Use

auvux login           # GitHub Device Flow, caches a JWT in ~/.auvux/
auvux whoami          # confirm

import auvux
from auvux import lanes, markers

# Same as `auvux login`, but inline. Idempotent — no-op if already
# signed in.
auvux.login()

canvas = auvux.Canvas("Beat Detection")
canvas.add(
    lanes.Waveform.from_path("demo.wav", size=160, label="mix"),
    markers=[markers.vlines(beats, color="#FFFF00")],
)
canvas.add(lanes.Heatmap(spect, yaxis="mel", fmin=30, fmax=11000, size=256))
canvas.add(lanes.line(t, beat_logits, color="#4CAF50", label="beats", fill=True))
canvas.minimap(from_lane=0)
canvas.clicks(beats, freq=800)
canvas.show()                  # opens at https://inference.auvux.com

Open https://inference.auvux.com in your browser before running canvas.show() — the publisher waits up to 60 seconds for at least one viewer to be present.

Waveform.from_path() accepts both local paths and http(s):// URLs; remote files are fetched once into memory and reused without a second round-trip on emit.

What gets sent

Each call to canvas.show() packages:

• A manifest.json describing the lanes and their layout
• Sidecar files: WAV / NPY / NPZ / CSV / JSON, per lane type
• Click-track time lists (if any)

The viewer pre-announces which file hashes it already has, so unchanged audio across iterations is sent once — subsequent runs reuse the cached blob.

Auth flow

auvux.login() runs the GitHub Device Flow with scope user:email:

1. Asks GitHub for a short user code (XXXX-XXXX)
2. Opens the browser to the verification URL
3. Polls until you authorise
4. Swaps the GitHub access token for an auvux JWT at https://inference.auvux.com/api/auth/cli-exchange
5. Caches the JWT in ~/.auvux/credentials.json (chmod 600)

Subsequent auvux.login() calls are no-ops if the cached token is still valid; pass force=True to re-authenticate.

Environment

Var	Default	Purpose
AUVUX_SIGNALLING_URL	from ~/.auvux/credentials.json, otherwise wss://signalling.auvux.com/signalling/me	Override the signalling endpoint (escape hatch for local dev)
AUVUX_API_BASE	https://inference.auvux.com	Auvux API host used for cli-exchange
AUVUX_GH_CLIENT_ID	built-in	GitHub OAuth App client id
AUVUX_PEER_WAIT	60	Seconds to wait for a viewer before erroring
AUVUX_ACK_WAIT	30	Seconds to wait per-viewer for delivery ack

Development

git clone https://github.com/auvux/auvux-py
cd auvux-py
pip install -e .

Release a new version: bump version in pyproject.toml, then

git tag v0.1.0a2
git push --tags

The release.yml workflow builds the wheel + sdist and publishes to PyPI via OIDC (Trusted Publisher) — no API token in the repo.

License

MIT — see LICENSE.