A drop-in replacement for W&B and Neptune
Open source experiment tracking that keeps up at scale. Dual-log from W&B or Neptune, validate, then switch.
5B+ metric points ingested, zero dropped
"Trainy removed the stress from neptune.ai being taken off the market. They stepped up, built Pluto, and raced to feature parity, and we are working with them as they build the first AI-native experiment tracker."
Geoffrey Négiar
Co-Founder @ The Forecasting Company
Safe migration with dual-logging, without touching your training code
Our compatibility layers for W&B and Neptune Scale let you dual-log to both platforms simultaneously so you can validate Pluto against real production usage before committing.
$ pip install pluto-ml
$ pluto login # or: export PLUTO_API_KEY=mlpi_xxx
import wandb # Your existing wandb code works unchanged
with wandb.init(project="my-project", config=config) as run:
# Training code here
run.log({"accuracy": 0.9, "loss": 0.1})
Install pluto-ml and authenticate. Dual-logging activates automatically: every wandb.log() call also logs to Pluto, with zero code changes.
Always open source
Never get locked out again. Self-host Pluto on your own servers with full source access. No vendor can sunset your experiment tracking.
Fork a run, branch a new experiment
Branch from any step of an existing run to explore training variations without starting from scratch. Try a new learning rate from a mid-training checkpoint, and stop burning GPUs on runs that will not converge.
import pluto
run = pluto.init(
name="lr-sweep-v2",
fork_run_id="MMP-42", # parent run
fork_step=500, # branch point
config={"lr": 0.001}, # override
)
for step in range(500, 1000):
run.log({"loss": 0.4})
forked from MMP-42 at step 500
> Compare train/loss across runs 100, 101, and 102. Which performed best?
pluto · compare_runs(metric="train/loss", runs=[100, 101, 102])
Run 102 has the lowest final loss
(0.118 vs 0.142 and 0.151) and converged ~2x faster.
Connect Pluto over the Model Context Protocol with one command, then ask in plain language.
Built for researchers who live in their dashboards
Explore an example project: every run logs 100k+ data points, including losses, gradient norms, and activations across layers.
Sub-second
Filter runs, compare metrics, and drill into any step in the live example without waiting, even with 100M+ data points in the project.
Every metric
Overlay losses, gradient norms, and activations across layers, then scroll and zoom into any step to see exactly where a run diverged.
Live
No signup and nothing to install. It is the real Pluto dashboard with real runs, open in your browser so you can click through every chart yourself.
Simple, predictable pricing
Start hosted for free, or self-host the open source. Pro adds generous caps with no metered surprises: no per-hour billing, no ingestion meters.
Pro
Hosted, managed
Up to 10 seats
10B data points/month
10 TB storage
Unlimited tracked hours
Run forking
Priority support
Enterprise
Larger teams
More than 10 seats
Custom data and storage limits
SSO and audit logs
Single-tenant or on-prem
Dedicated support and SLA
Free for academics. Sign up with your .edu email.
Sign upPrefer your own servers? Pluto is open source and free to self-host, and no vendor can ever sunset it.
Self-host guideMigrate to Pluto today.
5B+ metric points stored with zero dropped. Dual-log alongside W&B or Neptune, validate on real runs, then switch.
Live example with 100M+ data points