Skip to content

Quickstart

This walks through your first question, once Installation is done and the app is running at http://localhost:3000.

1. Upload data

Drag a CSV, TSV, Excel, JSON, NDJSON, Parquet, or Feather file onto the composer, or connect a database — see Connectors. Large files are sampled for analysis while the full file stays available in the workspace. Every loaded table becomes addressable to generated code as tables['name'], so cross-table questions need no extra setup step.

2. Ask a real question

Type it in plain language — "which customer segment had the highest churn last quarter, and why" rather than "compute churn by segment". The point of the agentic loop is that it's allowed to look before committing to an approach.

3. Watch it work

Every stage streams as it happens: the manager's reasoning, each move it makes and what it found, the generated code, the program's stdout, and the final answer token by token. If a step fails, you'll see the traceback and the retry — that's the self-correction loop, not a bug.

4. Pick a depth (optional)

Three depths are available in the composer:

  • Fast — one pass, no verification. Cheapest, least self-checking.
  • Auto (default) — the agent decides how much investigation the question needs.
  • Deep — investigate thoroughly, with a decision round-trip on every iteration and a verification pass at the end, regardless of model size.

5. Read the trust signals, not just the answer

  • Any number in the answer that doesn't trace back to real execution output is flagged — the agent doesn't get to invent a figure that looks plausible.
  • Silent decisions the code made (dropped nulls, an inner join, a top-N cut, a coerced date) are listed alongside the answer, because each one changes what the number means.
  • The headline result is independently recomputed by a different route; a mismatch is reported prominently rather than quietly resolved.

6. Keep the work

Every analysis is written out as a runnable script you can re-run next month against fresh data — see Exporting an Analysis. If you find yourself asking a structurally similar question again later, Wizard may offer to save it as a reusable skill.

Where next