The AI Data Scientist

From the scikit-learn company, a new teammate that ships machine learning models with the scientific rigor and statistical methodology baked in.
Any AI provider, any ML framework, any compute.

Schedule a demo Try our agent in demo mode

skore-agent · session 01 — probabl.ai/agent

skore-agent - methodology for agents, by the scikit-learn maintainers.

agent ▸ Watch me improve a real ML model, live. No install, no account. Pick a problem to start.

no problem loaded

try: pick churn · pick credit · pick salaries (or 1 / 2 / 3)

AI is transforming data science

Agents can draft a pipeline in seconds. The harder part, understanding, trusting and owning the result, still lands on you.

Automation, not options

You get a black box that runs, when what you wanted was visibility and choices. The agent decides; you lose the part of the work that made it yours.

The "why" goes missing

You can ship a result but not explain how the agent got there. Explainability is an afterthought in most stacks, and traceability matters more than raw output the moment a decision has to be defended.

Nothing reproduces

Hyperparameters, dataset versions and model lineage live in scattered places, so last week's experiment can't be rerun or compared with confidence. Iteration turns into guesswork.

A stack held together by hand

MLflow here, cloud storage there, a custom platform in between. Fragmented data across systems means costly cross-referencing before any real analysis can even start.

Skore is the rigor layer

Meet Skore

The data science agent that brings rigor to your ML workflow, right inside the tools you already use.

AI Agents

Your AI writes the code. Skore makes sure it holds up.

AI coding tools generate scikit-learn pipelines in seconds. Skore makes sure they're sound: it catches data leaks, picks the right metric for your problem, and flags the silent errors that only surface in production.

Automatic pipeline validation (structure, leakage, overfitting)
Smart metric recommendations for your use case
Works with Pi, Cursor, and ChatGPT-generated code

Skore Agent prompt box with model picker listing Claude, GPT, and Gemini

Setup

Two commands. Then it's just another model.

Skore lives in your harness's model picker and runs the experiment end to end: explore, build, evaluate, track. No new tool to learn.

Two commands to install: skore hub login, then skore agent init
Lives in your model picker, pick skore-agent like Claude or GPT
Works with opencode, Claude Code, and Pi, detected and pre-wired

Code editor showing skore-agent added to opencode.json with rigor enabled

No lock-in

Any agent. Any framework. Any compute. Fits your stack. Respects your data.

Skore is the layer between your team and your experiments, and it refuses to trap you in any one of them. Bring what you already use.

Any agent: Bring your coding harness. Claude Code, opencode, Pi, or any OpenAI-compatible client. Skore rides along as a model, not a migration.
Any ML framework: The scikit-learn world. scikit-learn, XGBoost, TabICL. The tabular stack behind 200M+ downloads a month, already native.
Any compute: Your cloud platform, or just your local environment with a pyproject.toml. Skore meets your stack where it already runs.

Diagram of harnesses, ML frameworks, and compute converging into the Skore Agent

Deployment options: Private cloud (BYOC), Your GCP Account, Your AWS Account, and SnowPark Container Service

Integrate it where the data lives

For your IT team that can't let data cross their boundary, Skore ships as a blueprint inside your own VPC: the agent integrates in your landscape, the rigor comes to you.

[Coming soon] Tabular Foundation Models for agents loop

For a decade, gradient boosting owned tabular data. That's changing. Foundation models learn from your table in a single pass, no training loop, no hyperparameter search. Built by the scikit-learn maintainers, they drop straight into the workflow you already know.

Stop trusting. Start verifying.

Install in two commands. Every pipeline your agent builds ships with the evidence attached, from the very first run.