remote-ssh-kaggle-vscode

# Remote-SSH Kaggle using Visual Studio Code **Connect to a Kaggle notebook over SSH from VS Code — password auth, no SSH keys required.** Keep a 12-hour session running uninterrupted, use a real terminal and debugger, and work with `.py` files instead of notebook cells.

---

Why

Kaggle’s notebook interface is limiting once a project outgrows a few cells. Connecting over SSH gives you the full VS Code experience against Kaggle’s GPUs: integrated terminal, breakpoint debugging, and a normal file-based project layout.

It also lets you stretch GPU quota. The default is 30 hours per week — if you stop the notebook session near the end of hour 29 and SSH back in, you get roughly 12 more hours, for about 42 hours a week.

Features

• 🔑 Password authentication — no SSH keypair to generate or upload
• 🧩 Modular bash scripts — easy to read, easy to customise
• ✨ Oh My Posh included — optional pretty terminal prompt
• ⚡ Quick setup — a handful of steps end to end

And plenty more to explore once you are in.

How it works

The Kaggle notebook clones this repository and runs three small scripts — see SCRIPTS_GUIDE.md for the details of each:

Script	Does
install_ssh_server.sh	Sets the root password and installs OpenSSH + ngrok
add_ngrok_token.sh	Registers your ngrok auth token
run_ssh_server.sh	Opens an ngrok TCP tunnel to port 22

---

Contents

• 1. Prerequisites
• 2. Set up the Kaggle notebook
• 3. Configure SSH in VS Code
• 4. Using it
• Tips and tricks
• Conclusion

---

1. Prerequisites

• Install Visual Studio Code: https://code.visualstudio.com/
• Create an Ngrok account: https://ngrok.com/

---

2. Set up the Kaggle notebook

• 2.1 Open the notebook: Notebook Example — or upload notebook_example.ipynb from this repository.

• 2.2 Choose Copy & Edit:

  

• 2.3 In the right-hand sidebar, pick one of these two GPUs:

  

  ⚠️ Warning: TPU is not supported.

• 2.4 Under persistence, select Files only so your files survive each Stop Session:

  

• 2.5 Go to Ngrok → Your Authtoken → press copy:

  

• 2.6 In cell 3 (the setup cell), set your SSH password and paste your Ngrok token:

    ssh_password = "kaggle"  # Change this to your desired password
  
    # Run bash scripts
    !bash install_ssh_server.sh $ssh_password
    !bash add_ngrok_token.sh YOUR_NGROK_TOKEN  # Replace with your actual token
  

• 2.7 In the last cell (cell 4 — runs bash run_ssh_server.sh), note the HostName and Port from the ngrok output, e.g. 0.tcp.ap.ngrok.io and 17520. You need both in step 3.6.

  

---

3. Configure SSH in VS Code

• 3.1 Press Ctrl+Shift+X, search for SSH, and install these two extensions:

  

• 3.2 For background on how Remote-SSH works, see the VS Code Remote-SSH docs.

• 3.3 Press Ctrl+Shift+P → Remote-SSH: Connect to Host…

  

• 3.4 Press Configure SSH Host…

  

• 3.5 Select ~/.ssh/config — usually the first entry in the list.

  

• 3.6 Add this block to the config file:

    Host Kaggle
        HostName 0.tcp.ap.ngrok.io
        Port 17520
        User root
  

  Field	Value	Where it comes from
  Host	Kaggle	Any name you like
  HostName	0.tcp.ap.ngrok.io	Step 2.7
  Port	17520	Step 2.7
  User	root	Always root

• 3.7 Press Ctrl+S, then Ctrl+Shift+P → Remote-SSH: Connect to Host…

  

• 3.8 Pick the host you just named — Kaggle:

  

• 3.9 When prompted, enter the password you set in step 2.6 (default: kaggle).

• 3.10 Press continue:

  

  💡 Tip: If VS Code asks you to choose the operating system, select linux.

• 3.11 The bottom-left corner confirms the connection:

  

---

4. Using it

• 4.1 Press Ctrl+K O, enter the path /kaggle, press ok.

  

• 4.2 Open a terminal with Ctrl+J. The system python already carries the full Kaggle stack (torch, numpy, …), so you can start working right away. uv is preinstalled for anything you need to add:

    uv pip install --system <package>
  

  Prefer an isolated environment? Create it with --system-site-packages so it can still see the preinstalled Kaggle packages:

    uv venv --system-site-packages .venv
    source .venv/bin/activate
    uv pip install <package>
  

  ⚠️ Warning: a plain uv venv, without --system-site-packages, starts empty — import torch fails inside it. And outside a virtual environment uv pip install needs the --system flag, otherwise it exits with No virtual environment found.

• 4.3 Activate CUDA. The image already ships the driver — it is simply missing from the SSH shell’s environment, because that shell does not inherit the notebook kernel’s paths. Two exports fix it instantly, with no download:

    export PATH=/opt/bin:$PATH
    export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:$LD_LIBRARY_PATH
  

  They apply to the current shell only — append them to /root/.bashrc if you open several terminals, and redo them after each Stop Session.

  📝 Fallback: if those paths move in a future Kaggle image, installing the driver utilities works too — it just downloads a few hundred MB, and also has to be repeated after every Stop Session: sudo apt install nvidia-utils-515 -y

• 4.4 Check the GPU is visible:

    nvidia-smi
  

  

  💡 Tip: If nvidia-smi reports no devices, the session has no GPU attached at all — check the Accelerator setting from step 2.3.

• 4.5 After each Stop Session, you only need to redo a subset:

  1. Run cell 4 to get the new hostname and port
  2. Update your SSH config with them (step 3.6)
  3. Reconnect from VS Code (steps 3.7 → 3.8 → 3.9)
  4. Carry on working (steps 4.1 → 4.2 → 4.3 → 4.4)

---

Tips and tricks

• To stretch GPU quota, stop the notebook session and SSH back in before you hit the 30-hour weekly limit — that gets you up to ~42 hours a week.
• Use the integrated terminal rather than notebook cells for anything shell-shaped.
• Set breakpoints and use the VS Code debugger instead of print debugging.
• Keep code in .py files and import across them, like a normal project.

Where your files live

The Data panel on the right has two sections, and they map to different paths with different rules:

Section	Path	Writable	Size limit
Input	/kaggle/input/...	❌ Read-only	~107 GB private, unlimited public
Output	/kaggle/working/...	✅ Your workspace	~20 GB

---

Conclusion

With Remote-SSH Kaggle and Visual Studio Code you get the full weight of Kaggle’s GPUs behind a development environment you actually enjoy using — a real terminal, a real debugger, and a normal project layout. Set it up once and the only thing you repeat between sessions is step 4.5.