Every browser asks you to make a trade. Chrome is fast — and funded by an ad business that needs your data. The new "AI browsers" (Comet, Atlas, Dia) bolt on a chatbot that ships your prompts to the cloud. MoltBrowser is built to remove the trade. How it compares:

a second account usually means a second browser profile, or logging out of the first. /sandbox reopens the page you're on in an off-the-record window: its own cookies and storage, gone on close. session isolation, not anonymity. the site and your network still see it.

someone sends you their details as one messy blob and you retype it into eight boxes. /paste reads your clipboard, probes the page for field labels (not the values in them, never password boxes) and splits the blob across them. the model doing the matching runs on your machine.

watching a page for a number usually means handing the url to a service that polls it for you. /watch does it on your machine: your browser opens the page on an interval, reads the text at a css selector, and notifies you with the value. it uses your own session, so logins work.

tab bankruptcy usually means closing everything and hoping nothing important was in there. /triage list prints what's open, close-inactive spares the pinned ones and the tab you're on, and closed tabs land in recently-closed. no model runs, so the list never becomes a prompt.

checking whether a site is sketchy usually means sending the url to someone else's server. /reputation asks your own machine instead: have you been here before, when, and how much you read there. it flags odd tlds and raw-ip hosts too, but that's a hint, not a verdict.

a recap of everything you read yesterday is a nice feature and a bad reason to hand someone your history. /digest reads the last 24h from the index on your own disk, groups it into a few themes and writes it up. on an on-device model, that list never leaves the machine.

sometimes you want a page's text and none of its scripts. /js off blocks javascript for that domain and its subdomains, reloads the tab, and remembers next launch. it's chromium's own per-site setting, surfaced as one command. plenty of sites break, and that's the trade.

filing a receipt usually means uploading it to someone's server. on an order page, /receipt reads merchant, date, total and line items and appends one row to a plain csv in your home folder. plaintext on purpose, so any spreadsheet opens it. the model runs on your machine.

click the same dead button four times in three seconds and moltbrowser notices. a small chip offers a clean reader view or a throwaway session, then vanishes after ten seconds. the detector is js inside the tab. it reports to nobody, so the signal stays yours.

chromium is a good browser wrapped in a lot of google. we strip that layer: the ai-mode "ask google" omnibox button is gone, the "you and google" settings pages are ours now, no telemetry to send. same engine, same extensions. the ai stays, it just runs on your machine.

you can't audit what you can't see. /trackers lists the third-party hosts the page actually loaded, grouped ads / analytics / cdn / other. it reads the page's own resource timings locally, so no urls leave the machine. the classifier is a short list, not easylist.

twenty tabs open and the answer is in one of them. /ask-tabs reads the text of every tab in the window, ranks the relevant bits, and answers from them. a cloud ai browser would have to upload all twenty pages first. on a local model, nothing leaves your disk.