ShellX Browser · built into shellX v0.3.4

A browser you and your
agent share.

ShellX Browser is a real, native browser — and the agent can drive it. Browse normally in your own profile, or hand a tab over and let it do the web work: research, sign-ins, form fills, checkouts. It reads and clicks real pages, fills credentials straight from your encrypted vault without ever seeing them, and logs a receipt for every step. It isn't a separate app — it ships inside shellX, wired to the Vault and the same approval gates.

Part of shellX · Windows, macOS & Linux · free & open source

ShellX Browser — user and agent web work in one desktop shell, with Personal / Agent Work / Disposable profiles, vault fill and secret capture, bookmarks, downloads and history, and observe / click / extract automation
01 What it does

Hand it a tab.
It does the web work.

Every one of these runs on a real page in a real browser — gated by your approval where it touches a secret, and logged so you can read back exactly what happened.

read & act

Read and act on any page.

Point it at a page and it reads the content back as clean, structured text, sees every element as a labelled reference, and clicks or types by reference — deterministic, not guessing at pixels or coordinates.

→ open · read · observe · click · type
wait for

Wait for the page.

Waits for an element or a condition before it acts, so it never clicks into a half-loaded page.

vault · fill

Sign in from your vault — blind.

Fills usernames, passwords and API keys straight from your encrypted vault. The value never appears in the chat, the logs, or the page source — the agent uses the secret without ever seeing it, and only after you grant the request.

→ vault grant required · value never revealed
profile card

Autofill a profile.

Name, address and card details fill from a saved profile card, with your approval — for checkouts and sign-ups.

email code

Read a one-time code.

Waits for the verification email and reads the one-time code, so a 2FA step doesn't stop the run.

$ agent wallet

Pay from a scoped wallet.

Completes a checkout from an agent wallet you cap and approve — it can spend up to your limit, and nothing more.

capture secret

Capture a secret to the vault.

Sees an API key or token on a page and deposits it straight into your vault — write-only, never read back into the chat.

screenshot · verify

Screenshot & verify.

Captures the page and checks it reached the state you expected before it moves on.

save · downloads

Save pages & downloads.

Saves a page, or handles a download as a real file on your machine — and hands you the path.

trace

Every action is a receipt.

Open Trace and replay every navigation, click, fill and capture the agent made — fully auditable, nothing hidden.

ShellX Browser · Personal profile · the agent-automation panel
The ShellX Browser open on a page, with the agent-automation side panel — autonomy control, Vault-ready badge, and Chat / Requests / Actions / Errors tabs
02 Share a tab

You drive, or
the agent drives.

It's a native browser, not a remote-controlled Chrome — pages load fast and the agent reads the real page. Set how much rope it gets per tab, watch its actions land in the side panel, and take the wheel back any time with Takeover or Abort. Hidden behind it: locks so two agents never fight over one tab, and redaction so secrets never leak into what the agent sees.

  • Per-tab autonomy — from "ask me first" to "go"
  • Live Chat / Requests / Actions / Errors view of what it's doing
  • Takeover or Abort at any moment — it's your browser
  • Drivable from your own code through the gated /browser API
Fill from Vault · a sign-in filled without the agent seeing the password
The Fill from Vault picker in the ShellX Browser, offering two matching saved passwords to fill a sign-in form, with the vault lock icon showing pending requests
03 The vault

It uses your secrets.
It never sees them.

This is why the browser isn't a separate app. It's wired to ShellX Vault: passwords, API keys, profile cards, email accounts and agent wallets live encrypted, and the browser fills them by name or a scoped grant. The value never crosses into the chat, the logs, or the page source — and every sensitive step goes through the Vault Request Center for your approval.

  • Approve, deny, review or revoke every request — secrets, profile fills, email codes, wallet spend, write-only deposits
  • Sign-ins and purchases come to you first; nothing happens silently
  • After a secret is filled, the agent can't even screenshot that page
  • Recovery kit, remembered-device unlock, and a password generator built in
04 Profiles

Three browsers,
cleanly separated.

Your everyday browsing, the agent's working browser, and one-shot throwaway tabs never share cookies, logins, or history.

01 / You Personal

Personal

Your own browser — your logins, bookmarks, and history. The agent doesn't touch it unless you hand it a tab.

02 / Agent Persistent

Agent Work

A persistent profile for the agent's own sessions — it keeps the logins and state it needs across runs, kept apart from your personal browsing.

03 / Throwaway Disposable

Task Disposable

A clean, throwaway tab for one task — no shared cookies, no history. When the job's done, it's gone.

05 Still a real browser

All the normal things, too.

It's a browser you'd actually use — the agent features sit on top, they don't replace the basics.

Bookmarks & history

  • Bookmarks with folders and a toolbar you arrange
  • Full history, per profile
  • Quick navigation and search from the address bar

Downloads

  • A normal downloads list and folder
  • The agent's saves and downloads land as real files, with paths

Privacy & ad-block

  • A privacy / ad-block mode you can switch on
  • Loopback-bound automation — nothing is exposed to the network
  • Per-profile separation keeps work and personal apart
06 Drive it from your code

One gated endpoint,
every browser action.

The same actions the agent uses are scriptable over a local, bearer-token-gated HTTP endpoint — the one surface, with redaction, locks, receipts and Vault grants all enforced on every call. There's no raw automation back door around the gates.

shellx.browser.action
01# token is written for you at ~/.shellx/shellxagent.token
02TOKEN=$(cat ~/.shellx/shellxagent.token)
03
04# open a page and read it back as clean text
05curl -X POST http://127.0.0.1:5757/browser/action \
06  -H "Authorization: Bearer $TOKEN" \
07  -d '{"action":"navigate","url":"https://example.com"}'
08curl ... -d '{"action":"extractMarkdown"}'
09
10# observe, then click a referenced element
11curl ... -d '{"action":"observe"}'
12curl ... -d '{"action":"clickRef","ref":"e12"}'
13receipt: logged · visible in Trace

full surface @ docs/API.md

It ships with shellX

Get the browser.
Get the whole app.

The Browser is part of shellX — install once and you get the agents, the builds, the vault, and the browser together. Free and open source.

signed · auto-updating · Windows 11 · macOS (Apple silicon) · v0.3.4 public beta