A terminal for your browser

lfl-terminal drops a command line onto any site. Known commands (search, open, ls, read, find, go, and more) run deterministically, no model involved. Type something else and a local model proposes exactly one action, which you approve, exact and literal, before it touches the page.

How it works

What it will not do

The honest limits are the pitch, not a footnote under it.

• A local model can still be argued into proposing the wrong-but-technically-allowed action. A page can try to bias which legitimate action it reaches for; this project does not claim otherwise.
• The human approval step depends on the human actually reading it. Whatever gets rubber-stamped still runs.
• Some things a page's own script can do (open a new tab, delay a redirect past the moment this extension is watching, a meta-refresh) are not always something this extension can see, let alone stop, even though everything it can see and block, it does.
• There is a public, itemized threat model on GitHub: what is covered, what is a disclosed residual, and why. Read it before trusting this with anything you actually care about.

It also plays snake

Alongside the commands above there is a small fun pack: snake and 2048 rendered as plain text frames right in the terminal panel, a few color themes, and a steam locomotive for anyone who mistypes ls. None of it is the point of the project, and it runs under the exact same limits as everything else here: no network calls, no touching the page, no exception.