Local-First · Zero Cloud · Open Source

Your experiments.
Your machine. Full control.

A lightweight, local-first ML experiment tracker built to run entirely on your own machine, keeping your metric logs private and offline by design.

Get Started View on PyPI
caliper_dashboard.app
Convergence Forecasting Zero Network Overhead JSONL Storage Step-Level Annotations Multi-Run Comparison GPU Monitoring 100% Offline MIT Licensed Convergence Forecasting Zero Network Overhead JSONL Storage Step-Level Annotations Multi-Run Comparison GPU Monitoring 100% Offline MIT Licensed

Everything you need.
Nothing you don't.

Fast, private, reproducible experiment tracking without the SaaS overhead.

Convergence Forecasting

Fits an exponential decay curve to live metrics. Predicts final validation values and flags plateaus, divergence, and oscillation automatically.

predicted_convergence: step 847 → val_loss 0.023
Loss Steps predicted

Step-Level Annotations

Pin text notes to exact training steps. Annotations render inline on chart plots so context stays with the data.

run.annotate(step=500, text="warmup complete")
warmup done LR decay

Multi-Run Comparison

Compare up to 8 runs side-by-side with overlaid loss and accuracy charts, hyperparameter diff tables, and best-of-run badges.

Dynamic charts · Overlaid comparison view
resnet50 vgg16 efficientnet

Project & Tag Filtering

Organize runs into projects with tag-based multi-select filtering. Shared tags surface runs across projects instantly.

caliper.init(project="CIFAR-10", tags=["baseline"])
CIFAR-10 baseline v2 ImageNet baseline finetune shared tag: baseline MNIST debug

Zero-Cloud Architecture

Metrics stream directly to local JSONL files in your workspace. Caliper operates entirely offline, requiring no API keys, accounts, or telemetry.

0 bytes uploaded · 0 accounts created
.caliper/runs/ Your Machine No cloud needed

Real-Time GPU Monitoring

Log GPU utilization, VRAM usage, temperature, and CPU load alongside training metrics. Spot bottlenecks per step.

run.log({"gpu_util": 94, "vram_gb": 11.2})
100% 50% 0% GPU VRAM CPU Temp 94% 82% 48% 62° Live

Up and running in under a minute.

Get Caliper integrated and running locally with three straightforward terminal and code steps.

1

Install

PyPI, Python 3.8+, no heavy dependencies.

terminal 
$ pip install caliper-py
2

Log

Two lines in your training script.

train.py 
import caliper

run = caliper.init(
  project="CIFAR-10"
)
run.log({"loss": loss})
run.finish()
3

Launch

Local server, auto-opens browser.

terminal 
$ caliper ui

Serving at http://localhost:5173
Scanning .caliper/runs/
Ctrl+C to stop.