Trackio
huggingface.co
- Free plan
- Windows
- macOS
- Linux
- Web
- API
- Self-hosted
What Trackio is
Trackio is a Python library for tracking machine-learning experiments, built on Hugging Face Buckets and Spaces. It records run metrics and configuration, then groups runs in a dashboard for comparison. The dashboard runs locally by default; logs can stay local, go to a Hugging Face Bucket, or be stored on a self-hosted server. Its logging API is compatible with wandb.init, wandb.log, and wandb.finish. It also accepts tables, Markdown reports, images, video, audio, 3D objects, HTML, and Matplotlib or Plotly figures. Optional packages support NVIDIA GPU and Apple Silicon monitoring, and documented integrations include Transformers, TRL, and RapidFire AI. Alerts can be stored, queried, displayed, or sent to Slack or Discord. The dashboard can expose an HTTP API and MCP server, with a write token required for mutation tools. Trackio is free. It does not include a first-class hyperparameter sweep API; its documentation describes using Python loops with a run for each configuration.
Who it suits
Trackio suits Python-based machine-learning teams and autonomous experiment workflows that need to record and compare runs. It may also fit users who want local or self-hosted log storage.
Good to know, and what to watch
- Free library and Hugging Face hosting.
- Dashboard groups runs for experiment comparison.
- Logs can include varied media and figures.
- Optional GPU and Apple Silicon monitoring.
- Integrates with Transformers and TRL.
- No first-class hyperparameter sweep API.
- Self-hosted ingestion and uploads require a write token.
Plans
Every plan and priceFree
Free plan
The verdict
Trackio covers experiment logging and run comparison, with flexible storage and a broad range of log types. Teams planning sweeps should account for the need to run configurations through Python loops.
What the maker publishes
- Alerts
- Alerts can be printed, stored, queried, shown in the dashboard, and optionally sent to Slack or Discord webhooks. source
- API and MCP
- The dashboard can run as an HTTP API and MCP server, exposing read tools and mutation tools gated by a write token. source
- Experiment data
- Runs can log metrics and configuration, and the dashboard groups runs to compare experiments. source
- Integrations
- Trackio documents native integrations with Transformers and TRL, plus a RapidFire AI integration for fine-tuning and RAG experiments. source
- Intended users
- Trackio describes itself as designed for autonomous ML experiments and provides CLI commands and Python APIs for agents to log and query experiment data. source
- Logging
- Its API is compatible with wandb.init, wandb.log, and wandb.finish, and the docs show it can be used as a drop-in replacement. source
- Media
- The logging API supports tables, Markdown reports, images, video, audio, 3D objects, HTML, and Matplotlib or Plotly figures. source
- Notable limitation
- Trackio intentionally does not include a first-class hyperparameter sweep API; its docs describe sweeps as Python loops with one run per configuration. source
- Purpose
- Trackio is a lightweight Python library for tracking machine learning experiments, built on Hugging Face Buckets and Spaces. source
- Self-hosting security
- For a self-hosted server outside Hugging Face Spaces, metric ingestion and uploads require a write token, which the docs say to treat as a secret on untrusted networks. source
- Storage and dashboard
- Trackio runs a dashboard locally by default and can store logs locally, in a Hugging Face Bucket, or on a self-hosted server. source
- System monitoring
- Optional packages enable NVIDIA GPU monitoring and Apple Silicon system monitoring, with compatible hardware metrics logged automatically by default. source
Checked against the maker's pages on 2 Oct 2026.