gradio-app/trackio

A lightweight, local-first, and free experiment tracking library from Hugging Face 🤗

Trackio – lightweight, free experiment‑tracking for humans and AI agents

What it is – Trackio is a Python library (with a Svelte‑based web UI) that lets you log the metrics, configs, artefacts and alerts of machine‑learning runs. It stores everything locally in a SQLite database (optionally frozen to Parquet) and can serve a dashboard on your machine or publish it to a free Hugging Face Space. The API is deliberately compatible with the popular wandb client, so you can drop‑in replace import wandb with import trackio as wandb.

Core features

  • Local‑first, cloud‑optional – logs stay on your disk by default; you can later sync to a Hugging Face Space for sharing.
  • SQLite storage + Parquet export – fast, high‑throughput writes via an in‑memory queue that flushes every 0.5 s.
  • CLI & SQL querying – trackio query … --sql "SELECT …" lets LLM‑driven agents read data directly.
  • WandB‑compatible API – trackio.init, trackio.log, trackio.finish work as a drop‑in replacement.
  • Dashboard – a Gradio‑styled UI (launch with trackio show) displays metrics, media, tables, alerts and can be embedded via an iframe.
  • Alerts & webhooks – send Slack/Discord/any‑URL notifications on custom conditions.
  • Free for unlimited personal/team use – no paid tier required.
  • Self‑hostable – run the backend/API on any server and point clients to it with TRACKIO_SERVER_URL.

Quick start

pip install trackio   # Python 3.10+

import trackio, random, time
trackio.init(project="demo", config={"epochs":5})
for epoch in range(5):
    trackio.log({"epoch": epoch,
                  "loss": random.random(),
                  "acc": random.random()})
    time.sleep(0.1)
trackio.finish()

Run trackio show (or trackio.show() in code) to open the dashboard.

Cloud & sharing

  • Hugging Face Spaces – pass space_id="user/space" to trackio.init and the library will automatically create or use a Space, handling authentication via huggingface-cli.
  • Sync – trackio.sync(project="my", space_id="user/space") uploads a local SQLite DB to a Space for collaborative viewing.
  • Embedding – use an iframe with query parameters (project=…, metrics=…, theme=dark, etc.) to embed a live dashboard in blogs or docs.

Performance

  • Non‑blocking log() calls; a background thread batches writes.
  • Local logging: effectively unlimited throughput (thousands of calls per second).
  • Space logging: ~2 k logs in < 8 s for a single run; ~32 k logs in ~14 s when 32 parallel threads are used (free‑tier Space hardware).

Development & extensibility

  • Backend written in pure Python; UI in Svelte 5 under trackio/frontend/.
  • Fork‑friendly: replace the frontend directory, rebuild, and the same API is served.
  • Documentation: https://huggingface.co/docs/trackio
  • License: MIT.

Who might use it?

  • Researchers who want a zero‑cost, offline‑first tracker.
  • LLM‑driven agents that need to log and later query experiment data via CLI/SQL.
  • Teams that prefer self‑hosting or Hugging Face Spaces over commercial services like Weights & Biases.

Pronunciation: TRACK‑yo ("track yo' experiments").

Related

  • Dispatch
    Hugging Face Trackio Release

    Hugging Face has released Trackio, a free, open-source, lightweight experiment tracking library that serves as a drop-in replacement for wandb with native integration for Hugging Face Spaces and Datasets.

  • Project
    jwuthri-cloud/Tracely-ai

    Tracely is a self‑hosted observability and CI/CD platform for AI agents. It ingests OTLP traces, evaluates runs with built‑in and LLM judges, clusters failures, and lets you promote a real production failure into a hermetic regression test that blocks pull‑requests. The stack (API, worker, UI, ClickHouse, Postgres, Redis, MinIO) can be deployed with Docker‑Compose or one‑click on Railway. The `tracelyai-sdk` auto‑captures LLM calls, and the GitHub Action integrates the gate into CI.

  • Project
    MigoXLab/dingo

    Dingo is an open‑source Python library (and SaaS product) for evaluating AI data, model outputs, and RAG systems. It supports rule‑based, LLM‑based, and agent‑driven checks, streams data from files, databases, S3, or Hugging Face, and provides JSON reports plus optional visual dashboards. Install via `pip install dingo-python` (with optional extras), run via CLI (`dingo eval …`) or Python SDK, and extend with custom rules, LLM prompts, or agents.

  • Project
    smallnest/pigo

    pigo is a Go‑based CLI AI coding assistant (a clone of the pi agent) that can read/write files, run shell commands, fetch web content, and interact with many LLM providers. It works in headless script mode or an interactive REPL, supports tool‑level allow/deny lists, project trust, prompt templates, skills, plugins, self‑update, and provides a Go SDK for embedding.

  • Project
    llt22/talkio

    Talkio is a cross‑platform (desktop + Android) chat client that lets you run multiple LLMs in a single group conversation, assign each a custom persona, and have them call external tools via the Model Context Protocol. It stores everything locally (SQLite, OS credential store) and supports rich rendering, file handling, Git integration, voice input, and more, built with Tauri 2 (Rust) and React 19.