lfl-terminal

About

lfl-terminal is a keyboard-first terminal that lives in your browser. Press the
backtick key (`) on any page and a floating terminal panel opens at your cursor
(drag its title bar to move it, pin it to keep it put). Type deterministic
commands to read, search, navigate, and act on the page. And for anything that
needs interpretation, a local AI model (one you run yourself) proposes a single
action that you approve or reject before it ever touches the page.

The whole point is local-first, human-supervised control:

- Local by default, no cloud AI, ever. Deterministic commands run entirely
  in your browser. The AI half talks only to the model endpoint YOU
  configure on loopback (127.0.0.1) - by default that's a model on your own
  computer, so nothing leaves it. A shared/cohort setup can point that same
  loopback address at a local bridge that relays to a cohort-operated model
  over a private tailnet instead - never the open internet, never to us -
  but no longer strictly on your machine. Either way: no accounts, no
  servers we operate, no telemetry.
- The model proposes; you decide. Any *model-proposed* action that changes
  the page (click, fill, navigate) is shown to you as a plain, deterministic
  summary and waits for your explicit approval. Typed commands you enter
  yourself (like `click <N>`) are your own direct intent and run without
  that card, under the same hard blocks. Read-only commands run without a
  gate.
- The model is never a security boundary. Approval, the fixed set of allowed
  actions, and the hard blocks (for example, credential fields are never
  auto-filled) are enforced by the extension's own code, not by the model.

What you can do:

- Navigate and act: go, open, back, click, fill, search, scroll
- Read and orient: read, ls, extract links/table, here
- Find and highlight: find, highlight (mark every match on the page), matches
  (list them with context)
- Make it yours: alias, macro, themes, per-site auto-open, a resizable and
  collapsible panel
- Ask the local model: prefix a request with "ask" (or just type a plain
  request) and approve the single action it proposes

Requirements (please read before installing):

- You need a local, OpenAI-compatible model server running at
  http://127.0.0.1:1238 (for example llama.cpp's llama-server, LM Studio, or
  Ollama behind a compatible endpoint). The deterministic commands work without
  it; the AI proposals need it.
- Chrome only, for now.

What it will NOT do (on purpose):

- It will not run shell commands, touch your filesystem, or use sudo. Ever.
- It will not send your page content or commands to any cloud service or to us.
- It will not let the model act on a page without your approval for anything
  that mutates it. (Commands you type yourself, like `click <N>`, are your
  own direct intent and execute immediately - that's you acting, not the
  model.)

Open source (Apache-2.0). Source, threat model, and issues:
https://github.com/MacadamiaButter/lfl-terminal