Skip to content
Get packs

Guide · Agent skill

How to use the Trackio experiment tracking agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

The huggingface-trackio skill teaches an AI agent to use Trackio, a library for recording machine learning training runs. After following this guide, you will know how to install the skill, ask your agent for logging, alerts and metric retrieval in plain language, and which privacy and safety settings to check first. The guide assumes you may be new to agent skills and explains each step.

Trackio Experiment Trackinghuggingface/skills · Apache-2.0

Quick start

  1. 1Download and unzip the skill. Download the huggingface-trackio ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder.
  2. 2Or upload it in the Claude apps. In the Claude apps, open Settings, then Capabilities, then Skills, and upload the skill.
  3. 3Install trackio. Run pip install trackio or uv pip install trackio in the Python environment your training uses.
  4. 4Ask for logging or alerts. Describe your training script and what to track or alert on, in plain language.
  5. 5Check privacy before syncing. If you sync to a Space, confirm the agent passes private=True when metrics should not be public.
  6. 6Read results as JSON. Ask the agent to use trackio list or trackio get with --json to inspect runs, metrics and alerts.

What the skill makes the agent do

A skill is a folder of instructions and reference files that an AI agent loads when your request matches its description. This one covers three tasks: logging metrics from Python training code, firing alerts for training problems, and retrieving logged metrics and alerts from the trackio command-line tool.

Trackio is a lightweight, free experiment tracking library from Hugging Face. It syncs to Hugging Face Spaces so you can watch a real-time dashboard, and it can send alerts to webhooks such as Slack or Discord.

  • Logging: the agent adds trackio.init(), trackio.log() and trackio.finish() calls to your training script.
  • Alerts: the agent inserts trackio.alert() calls for conditions like loss spikes, NaN gradients, low accuracy or stalled training.
  • Retrieval: the agent uses the trackio CLI, adding --json for machine-readable output.

Install the skill in Claude (ZIP or Settings)

Download the huggingface-trackio ZIP from this page and unzip it. For a personal setup, place the unzipped folder in ~/.claude/skills/. To limit it to one project, place it in that project's .claude/skills/ folder instead.

If you use the Claude apps, you can instead upload the skill under Settings, then Capabilities, then Skills.

  • Personal: ~/.claude/skills/
  • Single project: .claude/skills/ inside the project folder
  • Claude apps: Settings → Capabilities → Skills

Install the Trackio library

The skill does not replace the Trackio library. Your Python environment needs trackio installed, and the agent's code will import it. The source gives two install options, one with pip and one with uv.

Trackio's documentation is at huggingface.co/docs/trackio/index, and its source is on GitHub at gradio-app/trackio.

pip install trackio # or uv pip install trackio

Ask the agent for logging, alerts or retrieval

You do not need special syntax. Describe the task in normal language and mention what you need, for example that you have a training script and want its loss and accuracy tracked. The agent should match the request to the skill from its description.

For alerts, say which conditions matter to you, such as loss divergence or NaN values. For results, ask for the runs, metrics or alerts in JSON so you or another program can read them.

  • Add Trackio logging to my training script in train.py and log loss and accuracy each step.
  • Insert alerts for loss above 5 after 100 steps and for NaN values.
  • Use the trackio CLI to list alerts for my-project as JSON.

Understand privacy before syncing to a Space

Syncing to a Hugging Face Space is what makes metrics visible on a shared dashboard and lets them persist after a cloud instance shuts down. The skill notes that auto-created Spaces are public by default. If your metrics should not be public, the agent must pass private=True.

The skill's own minimal example uses space_id and private=True together. Check that setting in any code the agent writes before you run it.

  • Auto-created Spaces are public by default.
  • Pass private=True when metrics should stay private.
  • Syncing needs a Hugging Face account and token.
trackio.init(project="my-project", space_id="username/trackio", private=True)

Safety notes: scripts and network

This skill runs scripts and uses the network. The agent will run training scripts and trackio commands on your machine, and syncing or webhook alerts send data over the network. Review the code the agent writes and the commands it proposes, especially before launching long or background runs.

The skill is about instrumenting training scripts. It does not cover production monitoring of non-ML systems.

  • Read generated training code before running it.
  • Confirm space_id and private settings before any sync.
  • Be careful with webhook URLs, since alerts are sent to them.

What people use it for

  • Logging loss and accuracy from a training script into a shared live dashboard.
  • Adding alerts for loss spikes, NaN gradients or training stalls so an agent can notice problems during a run.
  • Letting an agent poll trackio list alerts with --json and then adjust hyperparameters for its next run.
  • Retrieving a run summary or metric values with trackio get for a report.

Frequently asked questions

Is the Trackio agent skill free?

Trackio is described in its source as a lightweight, free experiment tracking library from Hugging Face. The skill is published under the Apache-2.0 licence. Check the licence file in the repository for the full terms.

What do I need to use the huggingface-trackio skill?

You need Python with the trackio library installed and access to the trackio command-line tool. If you sync results to a Hugging Face Space, you also need a Hugging Face account and token.

Where do I put the skill folder?

Unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder for one project. In the Claude apps, you can upload it under Settings, then Capabilities, then Skills.

Are my Trackio metrics public?

They can be. Auto-created Hugging Face Spaces are public by default, so pass private=True when your metrics should not be public.

How is Trackio different from other experiment trackers?

The source describes Trackio as lightweight and local-first, with a wandb-compatible API for logging metrics. This skill adds agent-focused features such as alerts and JSON command output. This guide does not compare it with other tools on speed or features.

Can the agent watch a training run while it is running?

Yes. Alerts are printed to the terminal, so an agent watching the script output will see them. For background or detached runs, the agent can poll the trackio CLI for alerts instead.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use this skill if you are an ML engineer, researcher or TRL user who wants an AI agent to instrument training code, raise alerts and read machine-readable status. Skip it if you need private experiment data and have not planned for private=True, if you have no Python training code, or if you need production monitoring for non-ML systems.

Download the skill