- Python 3.10+ with pip
- Claude Code CLI installed (docs)
- Your data in CSV, DuckDB, or a supported warehouse
git clone <repo-url> ai-analyst
cd ai-analyst
pip install -e ".[dev]"claudeOn first launch, Claude will detect a fresh install and start the interactive setup interview. It walks you through:
- Your role and team: so Claude adapts its communication style
- Your data source: CSV directory, DuckDB file, or warehouse connection
- Your business context: what your company does, key metrics, team structure
- Your preferences: output formats, chart style, export channels
You can also run the setup manually at any time:
/setup
Once setup is complete, just ask a question:
What's our conversion rate by device type?
Or run the full analysis pipeline:
/run-pipeline
Place your CSV files in a directory (e.g., data/my_dataset/) and tell Claude
during setup. Each .csv file becomes a queryable table.
Point Claude to a .duckdb file during setup. DuckDB provides fast SQL
queries over local data.
Run /connect-data and follow the prompts for Postgres, BigQuery, Snowflake, Databricks,
Redshift, SQL Server, or MySQL. The analyst will select a native Python connector, CLI, API, or
MCP integration based on the source and available authentication. MCP is one interface option,
not a requirement for every warehouse.
To start fresh:
/setup reset
- Tier 1 reset: clears your profile and preferences
- Tier 2 reset: clears everything including dataset connections
python -m pytest tests/ -vai-analyst/
.claude/skills/ # Claude skill definitions and triggers
.knowledge/ # Context, definitions, corrections, and dataset manifests
agents/ # Pipeline-agent workflow definitions
helpers/ # Deterministic Python modules
tests/ # Pytest test suite
data/ # Public examples plus your gitignored datasets
docs/ # Documentation
outputs/ # Gitignored analysis outputs
working/ # Gitignored intermediate files and run records