Runicorn

Local-first experiment tracking for machine learning.
Python SDK, Web UI, Remote Viewer, and desktop workflow in one tool.

pip install -U runicorn

Python 3.10+ | Windows / Linux / macOS

What Runicorn is good at

100% Local Runs, logs, assets, and settings stay in your own storage directory.

Python-first Tracking Log metrics, text, images, datasets, pretrained weights, and code snapshots.

Modern Web UI Path tree navigation, compare mode, recycle bin, ZIP import/export, and theme presets.

Remote Viewer SSH-based remote access with saved connections, host key confirmation, and health monitoring.

Logging Compatibility Works with Python logging, torchvision MetricLogger, ImageNet meters, TensorBoard, and tensorboardX.

Desktop Option The current desktop build can open native windows for remote sessions on top of the same local viewer stack.

Install Runicorn, choose a storage root, create a first run, and open the viewer.

Connect to remote GPU servers with the current three-step wizard and OpenSSH-first workflow.

Learn the current page structure: experiments, assets, performance, remote sessions, and settings.

Go deeper into run lifecycle, asset tracking, compatibility helpers, and daily command-line workflows.

Remote Viewer became much more resilient: host key confirmation, faster environment probing, health monitoring, reconnect states, idle shutdown, and OpenSSH password support.
The Web UI moved toward a more productized layout: professional defaults, cleaner navigation, compare deep links, ZIP import/export preview, and a unified recycle-bin flow.
Logs and monitoring improved: virtualized log rendering, GPU history playback, and better chart/theme consistency.
Logging compatibility expanded beyond the original MetricLogger support to include ImageNet-style meters, TensorBoard, and tensorboardX patterns.
The desktop build now supports native remote-session windows instead of sending every session to an external browser.

Runicorn is open-source under the MIT License.