Experiment tracking for machine learning
Track every training run as files on disk and browse them in a local dashboard. Local tracking needs no server, database, or Docker. Sync to your own server or the hosted service preview to share runs with a team.
The PyPI package is 1.0.0, which tracks locally only and cannot sync to the self-hosted server or the hosted service:
pip install agenticml-liteVersion 3.0 adds workspaces and the hosted service to the self-hosted server. It is a pre-release, so install it from GitHub (the command follows the dev branch and changes as development continues):
pip install 'agenticml-lite[system] @ git+https://github.com/agenticML/agenticml-lite@dev'Three ways to use it
Section titled “Three ways to use it”LocalRuns are plain files on your machine. Nothing is sent anywhere. Start with the quickstart.
On-prem serverRun the 3.0 server on your own hardware, inside your network, for a shared dashboard and sync.
Hosted service (public preview)We run the server for you. Sign in at app.agenticml.xyz and sync your runs. Best effort, no SLA.
Why agenticml-lite
Section titled “Why agenticml-lite”Local-firstMetrics, config, code, and media are written to plain files under ~/.agenticml. Nothing leaves your machine unless you sync.
Zero-code trackingagenticml run train.py captures args, code, PyTorch components, and TensorBoard logs without editing your script.
Built-in dashboardagenticml ui serves charts, media, and per-run notes on localhost, straight from the files.
Fast and smallAbout 3 µs per log() call and about 8 bytes per scalar point on disk.
Start here
Section titled “Start here”InstallationInstall agenticml-lite and its optional extras.
QuickstartTrack your first run in under a minute.
Zero-code runsTrack an existing script with agenticml run.
Python SDKLog metrics, config, media, and artifacts from Python.
DashboardView runs, charts, and notes locally.
Distributed, resume, disabledMulti-GPU training, resuming runs, and turning tracking off.
Releases and status
Section titled “Releases and status”ReleasesWhat each version (1.0, 2.0, 2.1, 2.2, 3.0) includes and how to install it.
Service statusWhat is live on the hosted service and what is still open.