W&B-shaped experiment tracking that answers to SQL, not a SaaS login

What it does

Trackio is an experiment tracking library that mimics the wandb.init / wandb.log / wandb.finish API closely enough to be a drop-in replacement — alias the import and your existing logging code keeps working. Everything runs locally by default: metrics land in a SQLite database, and a Gradio-style dashboard (Svelte 5) visualizes metrics, media, tables, and alerts. If you want collaboration, it can deploy the same dashboard to a free Hugging Face Space instead.

The interesting bit

The real design target isn’t just humans — it’s agents. Because logs live in SQLite with a documented schema, an LLM can analyze your experiments by running read-only SQL through the CLI (trackio query --sql "SELECT ...") or the Python API. Your experiment tracker is now a database an agent can actually reason over, not a proprietary web UI it has to scrape.

Key highlights

• trackio.log() is non-blocking: it queues in memory and a background thread flushes to SQLite every 0.5 s, so the training loop never touches disk or network
• API-compatible with wandb — swap the import, keep the code
• Alerts can fire to webhooks (Slack, Discord, or any URL) and are queryable via CLI, Python, or HTTP
• Dashboards hosted on Spaces are embeddable as iframes with query params for project, metrics, zoom range, and theme
• Local projects can be synced to Spaces later; self-hosting the server over HTTP is also supported
• Logs can be “frozen” to Parquet for longer-term storage

Caveats

• Static Spaces snapshots are read-only and must be public — private dashboards require the Gradio Space flavor
• The README’s throughput section is truncated, so remote-logging rate limits (if any) aren’t visible here
• Requires Python 3.10+

Verdict

If you’re running ML experiments — especially with agents in the loop — and want tracking without an account or a cloud bill, this is a genuinely well-shaped tool. Teams already invested in wandb’s team features or heavy cloud integrations won’t find equivalents here, and that’s fine.