What it can do

Tools and capabilities

Search, the browser, documents, files and running code.

Web and the browser

search_web runs a query; read_url fetches a page and extracts its text. Both accept http and https on the public internet only, so a page saying "now read this local file" gets nowhere.

For sites that need clicking through, Draggy drives a real browser window, running uBlock Origin's filter lists on Ghostery's engine. Elements are handed to the model as an indexed list, so it clicks item 4 rather than guessing a selector.

A page behind a Cloudflare or captcha check is reported as needing a person rather than retried. Pass it yourself in the browser and the model can read the site afterwards: you share one session.

Making files

Word from Markdown, PowerPoint with a heading per slide, Excel from CSV, and PDFs typeset by the browser engine rather than dumped. Attach any of them back and Draggy reads them. A scanned PDF has no text to extract, and it says so instead of guessing.

Anything the system would execute on a double-click is refused: ask for .bat and you get told to use .txt.

Running code

run_code runs short Python and JavaScript in a scratch directory, stopped after twenty seconds. It is not a sandbox: it is a time limit, a folder and a stripped environment, and it runs as you.

Your own documents

Point Settings → Chat → Library at a folder and its contents are indexed on your machine, then searched by meaning and keywords at once. Structure helps: documents with headings chunk into meaningful passages. Index what you care about rather than everything.

Turning things off

Every tool can be disabled, and a disabled tool is not mentioned to the model at all. Fewer tools is fewer ways to go wrong, which is often the right trade on a small model.