kaggle-skill

Ask your coding agent about a Kaggle competition and get a short, sourced answer. Let it track where you stand, check a submission, and submit only after you say yes.

An independent, unofficial skill and plugin named kaggle, for Claude Code, Codex and other agents that load SKILL.md packages. Not affiliated with, endorsed by, or sponsored by Kaggle or Google.

You: Which Kaggle competitions with prize money are running right now? A short list with each one’s deadline, prize and metric, please.

A real session, recorded 2026-10-02 in Claude Code 2.1.286 (claude-opus-5-5): the agent ran competitions, then answered. The whole answer, cast.

What you need

• Python 3.11 or later, and an agent that loads skills.
• Nothing else to read public content: a competition’s facts and pages, solution writeups, discussions.
• For your own account (status, downloads, submissions, notebooks, publishing): a Kaggle credential, and two packages for the python3 your agent uses: python3 -m pip install "kaggle>=2.2.4" "kagglehub>=1.0.2".

Install

1. Add it to your agent.

   Where	Command
   Claude Code	/plugin marketplace add shepsci/kaggle-skill then /plugin install kaggle@shepsci
   Codex	codex plugin marketplace add shepsci/kaggle-skill --ref main then codex plugin add kaggle@shepsci
   skills.sh (Antigravity CLI, Cursor, and others)	npx skills add shepsci/kaggle-skill
   OpenClaw	clawhub install kaggle
   By hand	Clone this repository and copy skills/kaggle/ into your agent’s skills folder

2. Start a new session and ask something that mentions Kaggle: “What is the metric and deadline of the Kaggle Titanic competition?” The skill loads when Kaggle is named. It pre-approves only reading files, so in Claude Code’s default mode the agent asks before it runs each command.

3. For anything on your account, sign in once with kaggle auth login, or create a token with “Generate New Token” at kaggle.com/settings. Then ask “check my Kaggle setup”, which runs the skill’s doctor command. The setup guide has the details.

In Claude Code, shepsci/kaggle-skill is a marketplace that holds one plugin, which is why there are two commands.

What you can ask

• “Which Kaggle competitions with prize money are running, and what are their metrics?”
• “What is the metric, the deadline and the prize of this Kaggle competition?”
• “Summarize the rules that matter: team size, external data, submission limits.”
• “What did the top teams do? Preview their solution writeups.”
• “What are people discussing in this competition this week?”
• “Where do I stand: rank, submissions left today, best score, GPU hours?”
• “How far am I from the bronze line, and what moved since yesterday?”
• “Is this submission file well formed?”
• “Submit it.” (You get a dry run first, and it submits after your yes.)
• “Wait for the score and tell me how it compares with my validation.”
• “Download this dataset.” “Push this notebook and tell me when the run ends.”

What it does on its own, and what it asks first

KAGGLE_SKILL_READ_ONLY=1 in the environment makes every write refuse, and KAGGLE_SKILL_HIDE_ACCOUNT=1 leaves your entries and ranks out of listings and briefs, for screen sharing and recordings. Text that comes from Kaggle is marked as data, which lowers the risk that an instruction hidden in a forum post or a writeup is followed. The skill keeps a local record of what it submitted and how it scored in ./.kaggle-skill/; nothing is sent anywhere but to Kaggle.

How the pieces fit

An API token serves everything the skill does. An OAuth login (kaggle auth login) serves the Kaggle CLI and the reads on your account, not kagglehub: publish with --via cli. A legacy kaggle.json key serves the CLI and kagglehub, not the commands that read your standing.

See it work

You: What’s the deadline and the metric of kaggle.com/competitions/arc-prize-2026-arc-agi-3?

A real session, recorded 2026-10-02 in Claude Code 2.1.286 (claude-opus-5-5): the agent ran brief, then answered. The whole answer, cast.

You: What did the top three teams of Kaggle’s ARC Prize 2025 do to win? Two short lines each.

A real session, recorded 2026-10-03 in Claude Code 2.1.286 (claude-opus-5-5): the agent ran solutions and writeup, then answered. The whole answer, cast.

You: Submit ./submission.csv to the Titanic competition on Kaggle.

The agent checks the file, shows the dry run, and asks before it submits. A real session, recorded 2026-10-02 in Claude Code 2.1.286 (claude-opus-5-5): the agent ran submit, then answered. The whole answer, cast.

The sessions are recorded as they ran, read-only, and show nothing about an account; the demo library has the install demos and says how each one is made.

Update and uninstall

Run these in a shell, then start a new agent session to load the new version. Version 3.0.0 renamed the commands and made every write a dry run by default; the changelog lists what changed.

Documentation

Security, license, privacy

The skill does not sandbox the agent and does not limit which hosts it can reach. Review what an agent proposes before it submits or publishes. What the skill guarantees, and the tests that hold it to it, are in Architecture. Report security issues through SECURITY.md.

MIT license; see LICENSE. The copy on ClawHub is under MIT-0, as ClawHub requires. The skill collects no data: credentials and processing stay on your machine; see PRIVACY.md.