A Python library that brings Claude's data science expertise into your local workflow — Jupyter notebooks, scripts, terminal, anywhere Python runs.
No browser needed. No logging in. Just import and go.
pip install dataspark-ai # from PyPI
# or
pip install dataspark-ai[full] # includes sklearn, matplotlib, seaborn, plotly, scipy
# or from source
git clone https://github.com/KTG0409/dataspark.git
cd dataspark && pip install -e ".[dev]"export ANTHROPIC_API_KEY="sk-ant-api03-..."Get a key at console.anthropic.com/settings/keys
from dataspark import Spark
spark = Spark()
# Explore a dataset — get instant analysis, quality checks, and recommendations
spark.explore("sales_data.csv")
# Ask any data science question
spark.ask("Should I use one-hot encoding or target encoding for high-cardinality categoricals?")
# Design a complete project
spark.project("Build a customer churn prediction model for our SaaS platform")
# Brainstorm creative analysis ideas
spark.brainstorm("I have 3 years of e-commerce transaction data with 2M rows")
# Generate production code
spark.code("Build a feature engineering pipeline for time series with lag features and rolling stats")
# Get best practices guidance
spark.best_practices("Cross-validation strategies for time series data")
# Interactive session (like chatting with Claude)
spark.chat()Pass a CSV, Excel file, DataFrame, or URL. DataSpark will:
- Profile every column (types, distributions, outliers, correlations)
- Flag data quality issues
- Recommend specific analyses based on what it sees
- Ask you clarifying questions about your goals
- Provide ready-to-run code snippets
spark.explore("customers.csv")
spark.explore("https://data.example.com/dataset.csv")
spark.explore(my_dataframe, name="revenue")
# Focus on something specific
spark.explore("data.csv", focus="I need to predict the 'churned' column")Describe what you want to build. DataSpark designs the full pipeline:
spark.project("Forecast demand for 500 SKUs across 12 warehouses, daily granularity")
spark.project("Build a recommendation engine for our content platform")
spark.project("Anomaly detection for network traffic logs, ~10M events/day")Get creative, ranked ideas from quick wins to big bets:
spark.brainstorm("We have clickstream data, purchase history, and customer support tickets")
spark.brainstorm("Our marketing team wants to understand campaign attribution")Get complete, production-quality Python code:
spark.code("XGBoost pipeline with Optuna hyperparameter tuning")
spark.code("Automated EDA function that generates a PDF report")
spark.code("FastAPI endpoint that serves predictions from a pickled model")Maintains conversation history so you can have a back-and-forth:
spark.ask("What's the best way to handle class imbalance?")
spark.ask("Show me how to implement SMOTE with that approach")
spark.ask("Now how do I evaluate it properly?")Full interactive mode with slash commands:
/explore data.csv — Load and analyze a dataset
/project <desc> — Design a project
/brainstorm <ctx> — Generate ideas
/code <request> — Generate code
/model sonnet — Switch models
/save conversation.md — Save chat history
/clear — Reset context
/help — Show commands
/quit — Exit
# Model selection (default: Claude Sonnet)
spark = Spark(model="opus") # Most capable
spark = Spark(model="sonnet") # Balanced (default)
spark = Spark(model="haiku") # Fastest / cheapest
# Longer responses
spark = Spark(max_tokens=8192)
# Debug mode
spark = Spark(verbose=True)# Interactive chat
dataspark
# Explore a dataset
dataspark explore data.csv
dataspark explore data.csv -f "focus on the target variable"
# Quick question
dataspark ask "When should I use Ridge vs Lasso?"
# Project design
dataspark project "Build a fraud detection system"
# Use a specific model
dataspark -m opus explore big_dataset.parquetfrom dataspark import Spark
spark = Spark()
# Load data through spark — it profiles automatically
df = spark.load("data.csv")
# Now all questions are context-aware
spark.ask("What feature engineering should I do?")
spark.ask("Write the code for that")
# You can also explore at any point
spark.explore(focus="relationships between price and demand")dataspark/
├── __init__.py # Clean exports
├── core.py # Spark class — main interface & API calls
├── explorer.py # DataExplorer — load & profile datasets
├── profiles.py # DataProfile — statistical profiling
├── prompts.py # System prompts for each mode
└── cli.py # Command-line interface
The library works by:
- Profiling your data locally (pandas — nothing leaves your machine except the summary)
- Building rich context from the profile (statistics, distributions, quality issues)
- Sending that context + your question to Claude via the API
- Maintaining conversation history so follow-ups are contextual
Your raw data never leaves your machine. Only statistical summaries and column metadata are sent to the API.
from dataspark import Spark
spark = Spark()
spark._current_system = """You are a financial data science expert.
Focus on: regulatory compliance, risk modeling, backtesting.
Always consider: data leakage, survivorship bias, look-ahead bias."""
spark.ask("How should I backtest this trading strategy?")spark._data_context = """
We have a PostgreSQL database with:
- transactions (50M rows, 3 years)
- customers (2M rows)
- products (10K SKUs)
Business: B2B SaaS, $50M ARR, 15% annual churn
"""
spark.ask("What analyses would have the most business impact?")API costs per ~1000 tokens (approximate):
| Model | Input | Output |
|---|---|---|
| Haiku | $0.001 | $0.005 |
| Sonnet | $0.003 | $0.015 |
| Opus | $0.015 | $0.075 |
A typical explore() call uses ~2-4K tokens. An interactive session might use 10-50K tokens total.
Use spark = Spark(model="haiku") for cost-sensitive workloads.
- Your raw data stays local. Only statistical summaries (means, distributions, column names, 3 sample rows) are sent to the API.
- API key is yours. Use a personal key — it's billed to your Anthropic account, not tied to any employer.
-
No logging by default. Conversations are in-memory only unless you
/savethem. -
Review what's sent. Call
spark.explorer.context_for_llm()to see exactly what goes to the API.
MIT — use it however you want.
