Epokio

Watch, compare, queue and report machine learning training runs from a web page, a tray icon, a terminal or the macOS menu bar. Epokio reads the files your framework already writes, so there is no logging code to add, no account and no cloud.

What it does

• See every run on this machine or a remote GPU box: state, epoch, time left, best score, curves, plain-language notes ("recall is much higher than precision…") and the result images your framework saved.
• Get told when a run finishes, fails (NaN loss), stalls or reaches a target score: on the Mac, in the tray, or as a phone push (ntfy, Slack, Discord, Telegram).
• Compare runs: only the settings that differ, a settings table for a whole sweep next to each run's main score, CSV export and Markdown reports.
• Start and queue runs (Ultralytics YOLO) from the web page, one at a time on each GPU, with a data check before starting, a one-button Python setup (CUDA PyTorch on NVIDIA machines), resume from weights/last.pt, and failure causes in plain words.
• AI assistants can read and queue runs through the MCP server (pip install "epokio[mcp]").

Reads Ultralytics (results.csv), Hugging Face Trainer (trainer_state.json), PyTorch Lightning (metrics.csv), Keras (CSVLogger) and TensorBoard event files (read without TensorFlow). A hand-written loop shows up with two lines:

import epokio
with epokio.start("runs/my-model", epochs=50, lr=1e-4) as run:
    for epoch in range(50):
        ...
        run.log(val_loss=vl, accuracy=acc)

Start

pip install epokio
epokio setup            # finds your runs, starts the helper, opens the page
epokio setup --lan      # also let your phone or another computer watch this machine

Other commands: epokio watch (terminal view, good over SSH), epokio tray (pip install "epokio[tray]"), epokio doctor (what Epokio sees, for a bug report), epokio agent --stop.

Viewing is open on this machine; anything that starts, stops or changes something needs the machine's token (epokio agent --show-token). The helper uses only the Python standard library.

MIT licence.