Hugging Face turns spreadsheets into an AI playground

AI Sheets is a new open-source, no-code tool from Hugging Face that lets you build, enrich, and test datasets using open AI models inside a familiar spreadsheet.

What it does

Instead of writing Python scripts to batch-process data through APIs, AI Sheets turns prompt engineering into spreadsheet columns.

You can start by importing an existing dataset in CSV, TSV, XLS, or Parquet format with up to 1,000 rows. If you don't have data, you can simply type a prompt to generate synthetic data from scratch, like asking for a list of fictional startups with their industry and tagline. It generates five sample rows, and you can drag down the column to generate more rows instantly, including images.

Adding new columns is where it shines. You hit the plus button, pick a model, and write a prompt referencing existing cells, such as summarizing text or classifying sentiment. There is also a web search toggle if your prompt needs fresh factual data, like looking up zip codes for addresses.

Its best feature is how it handles prompt tuning. When an AI cell produces something slightly off, you don't need to rewrite your system prompt. You just edit the cell directly or hit the thumbs-up button on a good result. Under the hood, AI Sheets automatically pulls those edited and liked cells in as few-shot examples for subsequent generations.

You can also use it for vibe testing. You can set up two columns to run the same prompt across different models, such as Qwen and gpt-oss, and then add a third column with an LLM-as-a-judge prompt to automatically evaluate which output is better.

How to run it

Getting started takes zero setup if you just want to experiment. Hugging Face hosts a free web version on Spaces at huggingface.co/spaces/aisheets/sheets.

For local workflows or private datasets, the project is completely open source under huggingface/sheets on GitHub. It connects to thousands of open weights on the Hugging Face Hub through inference providers, and it can also hook into locally hosted models.

Once your dataset looks good, you can export it back to the Hugging Face Hub. The export bundles a YAML configuration file containing your exact prompts and few-shot examples. When you need to scale beyond the 1,000-row spreadsheet limit, you can pass that config directly into a HF Jobs script to process full-scale production datasets in the background.

Why it matters

Batching LLM tasks usually forces you to pick between clunky API scripts and heavy automation workflows. AI Sheets bridges the gap by wrapping model experimentation in a UI that anyone who knows Excel already understands.

The feedback loop where in-cell edits instantly turn into few-shot context is genuinely clever and saves massive amounts of prompt-crafting time. While the in-browser limits mean you won't process millions of rows directly in the sheet, it is easily one of the smoothest tools available for quick data cleaning, synthetic prototyping, and blind model evaluations.