---

Why?

Kaggle gives you free GPUs (2x T4, 30h/week) and free TPUs (v3-8, 20h/week). But you're stuck in a browser notebook. kgz lets you drive Kaggle from your terminal, scripts, or AI agents.

from kgz import Kernel

k = Kernel("https://kkb-production.jupyter-proxy.kaggle.net/k/.../proxy")
result = k.execute("import jax; print(jax.devices())")
# [CudaDevice(id=0), CudaDevice(id=1)]

Install

pip install kgz

Single dependency: websocket-client.

Getting Your URL

1. Open a notebook on kaggle.com/code
2. Enable GPU or TPU in Settings
3. Copy the URL from your browser bar
4. Pass it to kgz

Features

Execute with Structured Output

result = k.execute("print('hello'); 2+2", stream=False)
result.success        # True
result.stdout         # "hello\n"
result.return_value   # "4"
result.elapsed_seconds # 0.1

GPU & TPU Support

k.is_tpu()            # True if TPU kernel
k.device_info()       # Full device details
k.tpu_type()          # "TPU v3-8" or "Tesla T4"
k.resources()         # GPU util%, VRAM, CPU%, RAM

Health Dashboard

k.health_check()
#   Kernel:  busy
#   Backend: gpu (2 devices)
#   GPU 0:   85% util, 12000/15360 MB
#   GPU 1:   82% util, 11500/15360 MB
#   CPU:     17%
#   RAM:     8.6/33.7 GB
#   Train:   step 1234/5000 | loss 2.31 | 56,000 tok/s
#   ETA:     ~35m
#   Quota:   27.9h remaining (GPU)

Quota Tracking (30h/week GPU, 20h/week TPU)

k.start_quota_tracking()
# ... train for hours ...
k.stop_quota_tracking()
k.quota_summary()
# GPU: 4.2h used / 30h quota (25.8h remaining)
# Session: 7.8h before expiry

Output Caching (98x speedup)

result = k.execute_cached("pip install -q jax flax")
# First: 5s (remote), Second: 0.001s (local cache)

Pipeline (Run-Once)

results = k.pipeline([
    ("Install", "pip install -q jax flax"),
    ("Check GPU", "import jax; print(jax.devices())"),
    ("Train", "train(steps=5000)"),
], notify_url="https://hooks.slack.com/...", use_cache=True)
# Tracks quota, caches setup, notifies Slack on complete/fail

Notebook Import & Export

results = k.run_notebook("training.ipynb")   # Run existing .ipynb
k.to_notebook("output.ipynb")                # Export history as .ipynb

Kaggle Datasets

k.list_datasets()                        # Show mounted datasets
k.attach_dataset("openai/gsm8k")        # Check availability

File Operations

from kgz import upload_file, download_file, FileSync

upload_file(url, "model.py", "model.py")
k.download_model("/kaggle/working/model.pkl", "./model.pkl")

sync = FileSync(url, "./src", "/kaggle/working/src")
sync.start()   # Background watch — auto-upload on change

Secrets (excluded from history & export)

k.set_env(WANDB_API_KEY="...", HF_TOKEN="...")

Environment Snapshot

k.snapshot_env()    # pip freeze → local file
k.restore_env()     # Restore packages (after kernel restart)

Budget Alerts

k.set_budget(max_hours=8, notify_url="https://hooks.slack.com/...")
# Alert at 80%, interrupt at limit

Sessions & Profiles

k.save_session()                          # Save for later
k = Kernel.resume("my-session")          # Resume

k.save_profile("gpu-training")           # Save config
k = Kernel.from_profile("gpu-training")  # Reuse

Notifications

k.execute_notify(code, notify_url="https://hooks.slack.com/...", label="Training")

Parallel Execution

k1, k2 = Kernel(url1), Kernel(url2)
results = Kernel.parallel_execute([k1, k2], "import jax; print(jax.devices())")

Persistent Connection

WebSocket reused across calls — no reconnect overhead:

for i in range(100):
    k.execute(f"step({i})", stream=False)  # ~100ms each, not 500ms

Error Handling

from kgz import KernelError
try:
    k.execute("1/0", raise_on_error=True, stream=False)
except KernelError as e:
    print(e.result.traceback)

CLI

kgz run URL "code"              # Execute code
kgz exec URL -f script.py       # Execute file
kgz status URL                  # idle/busy
kgz interrupt URL               # Ctrl-C
kgz wait URL                    # Block until idle
kgz upload URL file [remote]    # Upload
kgz download URL remote [local] # Download
kgz ls URL [path]               # List files
kgz snapshot URL                # Variable inspection
kgz resources URL               # GPU/TPU/CPU usage
kgz sync URL local_dir          # Watch & sync
kgz notebook URL -f cells.txt   # Run notebook cells
kgz sessions                    # List saved sessions

Kaggle Limits

Resource	Weekly Limit	Session Max
GPU (T4)	30 hours	12 hours
TPU (v3-8)	20 hours	9 hours

Use k.quota_summary() to check remaining time.

Documentation

• Usage Guide — Complete feature walkthrough
• Agent Guide — Patterns for AI coding agents

Pair with tpuz

pip install kgz     # Kaggle free GPUs/TPUs
pip install tpuz    # GCP TPU/GPU pods

Claude Code

mkdir -p ~/.claude/skills/kgz-guide
cp SKILL.md ~/.claude/skills/kgz-guide/skill.md

License

MIT