Pretzel AI

Start typing and a suggestion lands after roughly a second. Tab takes it. The default engine is Mistral's Codestral, a fast model tuned for code completion, and settings let you point it elsewhere. This is the feature you notice first because it fires without a keystroke.

Cmd+K or Ctrl+K opens a prompt box over the current cell. Existing code gets edited in place, highlighted code limits the scope, and a prompt starting with inject or ij asks for new lines instead of a rewrite. Typing @ pulls up variables in memory, so you can reference a real dataframe by name.

Ctrl+Cmd+B opens a chat panel that sees the active cell plus relevant code from the rest of the notebook. Ask it to explain a block, find the function that drops outliers, or write a query against a frame. Answers come back as code you can insert.

When a cell raises an exception, a Fix Error with AI button appears in the corner of the output. Click it and Pretzel AI reads the traceback along with the cell, then suggests a correction. You keep the final say on whether to apply it.

This is a fork of Jupyter, not a wrapper around it. Existing .ipynb files open unchanged, and Jupyter config, settings, keybindings and extensions work out of the box. A team that already runs JupyterLab can test it on one machine without a migration.

The hosted server needs no configuration, but settings let you route chat and completion to OpenAI, Anthropic, Groq, Azure or Ollama with your own keys. Requests then travel from your machine to the provider. That path is the reason local models matter here for anyone who can't send code to a third party.

The browser version at pretzelai.app is free and creates an account from your email address. The pip package costs nothing too, and a Dockerfile covers Windows setups where build tools get in the way. Both routes reach the same interface.