Probe vs Trackio
Trackio is a free, lightweight experiment tracking library from Hugging Face. Its API matches wandb.init, wandb.log and wandb.finish, its dashboard runs locally by default, and it can log to a Hugging Face Space or a server you host. Probe is a hosted record of a whole team's ML work, built for the coding agents that run it.
Tell your coding agent:
Install Probe on this machine and its coding agents. Read https://research.prbe.ai/install.md and follow it.
How they differ
| Probe | Trackio | |
|---|---|---|
| What it is | A hosted team record: projects, experiments, runs, notes and the agent sessions behind them | A lightweight Python tracking library, under 3,000 lines by its own description |
| Where data lives | Probe's hosted service | Locally by default, or a Hugging Face Space or bucket, or your own Trackio server |
| Logging from code | probe.init() and probe.log() | import trackio as wandb, then the usual calls |
| Coding agents | Agents sign you up and install Probe; work is recorded as it happens, and agents read the team's record with probe ask or a hosted read-only MCP server | CLI commands and Python APIs designed so LLMs can log and query experiment data, plus an MCP server mounted alongside the dashboard with read tools for projects, runs and metrics |
Trackio facts above are from the Trackio documentation and its page on Trackio as an API and MCP server, checked November 2026.
Pick Trackio if
- You want a free, minimal tracker that runs on your laptop.
- You already live on Hugging Face and want dashboards on Spaces.
- You want to fork a small codebase and add what you need.
Pick Probe if
- Several people and agents work on the same research and need one shared, hosted record.
- You want each result linked to the agent session and code that produced it.
- You want your agent to set the tracker up for you: sign-up, install and wiring into your scripts.
- You track RL rollouts and want them linked to runs and the agent sessions behind them.