signin-window

Human sign-in windows, reusable session-maintenance primitives and optional bounded browser-challenge mechanics for self-hosted applications.

Let a person sign in to a website themselves, from their own browser or phone, in a browser window that runs on your server. Your server keeps the session afterwards; the password, the one-time code and the site's checks stay with the person.

It is made for MCP servers and other headless services that act on a site for someone (shops, rides, photos, mail) but should never hold their password or automate a sign-in. Your assistant calls a tool, gets a one-time link, and gives it to the person. They open the site's own sign-in page through that link, and the page notices when they are in.

person's browser ──https──> your server (relay) ──private network──> login service
  the link's page              /connect/<id>                          Xvfb + x11vnc + Chromium
  noVNC viewer  <──────────── live view (WebSocket) ────────────────  on the site's sign-in page

• The login service (signin_window.service) runs sign-in windows: a virtual display, a remote-desktop server on loopback, and your runner, which opens the site's sign-in page on that display (with Playwright, say) and waits until the site has signed the person in. It stays on a private network, behind a bearer.
• The relay (signin_window.relay) runs in your server: one-time links, their page, noVNC's files and the live view, carried through to the login service. When the window ends, your hook stores the session.
• The paste add-on lets the person paste into the remote page with Ctrl+V (Cmd+V), which noVNC alone does not.

The human window and relay APIs never click, type or fill a page: the person does every step of the sign-in. They never import optional automation modules.

Applications can also use the session-maintenance primitives and explicitly configured challenge mechanics described below. These are separate capabilities: enabling a human sign-in link never enables automated interaction.

Session maintenance

signin_window.sessions supplies durable attempt accounting, warm-first upkeep and refresh scheduling without choosing a site's session format or policy. It starts no worker and reads no credentials. Applications own account checks, holds, schedules, session verification and publication.

from pathlib import Path
from signin_window.sessions import AttemptBudget, WarmFirst

budget = AttemptBudget(Path("/private/state/example-account/attempts"))

def sign_in_once():
    if not budget.reserve("example", limit=1):
        raise RuntimeError("sign-in budget exhausted")
    # Call the application's explicitly authorized sign-in adapter here.

# Supply application callbacks and its positive signed-out exception:
# flow = WarmFirst(refresh_and_verify, sign_in_once, signed_out=SignedOut)
# published_generation = flow.run()

WarmFirst calls refresh first. Only the supplied signed-out exception permits one sign-in and one subsequent refresh; transient failures, challenges and inconclusive evidence propagate without signing in. SessionStillDead means the new session remained signed out; the application should retain an operator hold. refresh_due coordinates regular checks, failure spacing and bounded early-refresh requests. No helper treats a successful sign-in as entitlement to application content.

AttemptBudget records a reservation before dispatch under a process-safe lock. It retains interrupted and failed attempts. The default rolling window is 24 hours; files use {"attempts": ["2030-01-01T00:00:00+00:00"]}. Malformed or symlinked state fails closed. Keep each tenant/account's root separate and preserve its ledger during upgrades. signin_window.state provides atomic private writes and bounded regular-file reads; signin_window.leases provides ordered session, attempt and profile leases. All processes sharing a profile must use the same lease path.

Optional challenge automation

Install signin-window[automation] for typed visual plans and bounded hCaptcha and reCAPTCHA image-grid mechanics. Importing these modules starts no browser or model request. The caller supplies the browser, a VisionPlanner, explicit authorization and site-specific acceptance checks. There is no built-in model account, credential store or implicit planner fallback.

• signin_window.automation.hcaptcha.HCaptchaVisualMechanic accepts an explicit planner and acceptance probe; only independently observed acceptance succeeds.
• signin_window.automation.recaptcha.RecaptchaImageSelectSolver handles a visible image grid; its result only says the challenge cleared.
• signin_window.automation.plans validates bounded grid, point and drag plans. A slow planner cannot apply an action to a replaced challenge task.
• signin_window.automation.detection.detect_challenge passively classifies challenge surfaces. automation.resolver.ChallengeResolver dispatches only the caller's configured mechanics, with explicit mode and deadline. A cleared challenge still needs the application's independent access check.

Keep human and automated runners separate. Credential entry, CAPTCHA handling, and application-content verification require their own explicit policy. The mechanics never establish account identity or content access themselves. Private diagnostic images and session state must remain outside logs and repositories.

Profile preparation and shutdown helpers in signin_window.automation.profiles refuse to steal a live or foreign browser's profile. OwnedBrowserProcess captures an owned child through a Linux pidfd, so cleanup cannot signal a reused PID. Hold the shared profile lease before preparation and through browser exit. These helpers and native pointer input require Linux; importing the base window and relay APIs does not import them.

When adopting an existing maintenance system, retain its session formats, attempt-ledger files, holds and profile paths. Upgrade all processes sharing a profile together if their lease naming changes. Rehearse with synthetic state and test successful reuse, interrupted work, stale candidates and concurrent writers. No helper automatically acknowledges a hold or resets an attempt budget.

Install

pip install signin-window

Python 3.11 or later. The login service runs on Linux with Xvfb, x11vnc, noVNC 1.5 or later (its static files) and a browser; examples/minimal/Dockerfile builds one. Use Debian trixie's Xvfb and x11vnc (CI runs the end-to-end check there): on Ubuntu 24.04's, the live view showed the window but no click or key reached it.

The login service

A runner gets a Window. It opens the sign-in page on the window's display, calls window.waiting(host) once the page is up (the live view opens only then), and returns the session when the site has signed the person in:

import asyncio

from playwright.async_api import async_playwright
from signin_window.browser import LAUNCH_ARGS, harden
from signin_window.hosts import allowed, host_resolver_rules
from signin_window.service import Windows, build_app

HOSTS = ("example.com",)

async def sign_in(window):
    async with async_playwright() as p:
        context = await p.chromium.launch_persistent_context(
            window.temp_profile(), headless=False, env=window.env(), no_viewport=True,
            args=[f"--host-resolver-rules={host_resolver_rules(HOSTS)}", "--window-position=0,0", *LAUNCH_ARGS],
        )
        await harden(context)
        await context.route("**/*", lambda r: r.continue_() if allowed(r.request.url, HOSTS) else r.abort())
        page = await context.new_page()
        await page.goto("https://www.example.com/signin")
        window.waiting("example.com")
        while not window.cancelled.is_set():
            cookies = await context.cookies()
            if any(c["name"] == "session" for c in cookies):
                return {"cookies": cookies}          # handed over to the server, once
            await asyncio.sleep(1)

app = build_app(Windows(sign_in), token=TOKEN, novnc_dir="/opt/novnc")

examples/minimal/login_service.py is the complete version.

Locking the browser down

The window is a real browser in front of a person: signin_window.browser makes it show the site's sign-in and nothing else.

• LAUNCH_ARGS (kiosk mode): no address bar, tabs, menus or window frame, and every window fills the display (which has no window manager). DevTools stay closed.
• POLICY, Chromium's managed policy: the image installs it as root (python -m signin_window.browser install-policy; check-policy tests for it, and policy_installed() lets a service refuse to open a window without it). Chromium does not offer to save passwords or fill addresses and cards, and opens no local files, file dialogs, downloads, printing, browser sign-in or sync; its own pages (new tab, history, settings) and file:, data: and view-source: addresses are blocked.
• harden(context) (harden_sync for Playwright's sync API): closes a tab or window opened with Ctrl+T or Ctrl+N. Kiosk mode reaches the windows of a persistent context only: for browser.new_context(), call harden(context, browser), which makes its windows full screen (kiosk mode keeps them so) and closes windows opened outside it.

None of this changes what a page sees of the browser, so a site may still offer a passkey, which cannot work in a remote window. Removing WebAuthn from the pages would hide the offer, but a Chrome without it contradicts itself where sign-in pages' bot checks look; DevTools' WebAuthn environment keeps it but makes sites show a passkey button instead.

scripts/browser_check.py checks all of it in a real Chromium; CI runs it on every change. In your route check, tell the page's own address from what it loads with hosts.top_level(request): Playwright has no frame yet for a popup's first page, and asking for it raises.

A window ends in one of two ways:

• Hand-over: the runner returns the session. It waits up to five minutes to be taken, once, by the server, then it is dropped. The login service keeps nothing.
• Kept: the runner keeps the session itself (a browser profile it refreshes, a snapshot it writes for the server to read) and calls window.finish("connected"), or another final state of its own such as "wrong_account". window.report(account=...) adds public fields to the window's state. Once the person is signed in, call window.saving() while you keep the session: the live view closes and closing the window no longer interrupts you. Pass a lock to Windows if a keeper uses the same profile, and windows.close_key(key) closes every window of an account you remove.

A runner can be a plain function too (it runs in a thread and should watch window.cancelled), or a command: signin_window.child.run_child(window, argv) runs it on the window's display and follows its progress lines (SIGNIN {"state": "waiting"} ... SIGNIN {"state": "signed_in", ...}), which suits sign-in code that is synchronous or someone else's.

Windows options: slots (windows at a time, each with its own display), minutes (time limit), xvfb and x11vnc (binaries), lock and lock_wait, display_base and vnc_base (window n uses display display_base + n and VNC port vnc_base + n on loopback). build_app options: start_params(body) turns a start request into the window's params (raise ValueError to refuse it; a "key" param makes a newer window for the same key cancel the older), title for the plain view page, and routes of your own behind the bearer.

The service's API (build_app): POST /logins, GET /logins/{id}, POST /logins/{id}/result (the session, once), DELETE /logins/{id}, WS /logins/{id}/vnc?t= (the live view, with the window's view token), GET /view/{id}?t= (a plain page showing the window, for someone on the service's own network), GET /novnc/... and GET /health. Shutting the service down closes its windows. LoginClient.request(method, path) calls a route of your own on it.

The relay, in your server

from signin_window.client import LoginClient
from signin_window.page import PageText
from signin_window.relay import Relay

def store(link, session):
    save_session(link.key, session["cookies"])  # yours
    return "connected"

relay = Relay(
    LoginClient("http://login:8080", TOKEN),
    "https://mcp.example.com/connect",
    text=PageText(service="Example", secure_host=r"(^|\.)example\.com$"),
    on_signed_in=store,
)
app.routes.extend(relay.routes("/connect"))   # outside your server's bearer

link = relay.create(account_id)               # in your connect tool
return {"url": link.url, "expires_at": link.expires_at}

A link works for link_ttl (20 minutes) and one successful sign-in, up to max_starts windows (a reload or "try again" opens a new one); a newer link for the same key replaces it, and each key gets links_per_hour links. Its id has 192 random bits; requests for ids that do not exist are counted per client address (from X-Forwarded-For, or your client_address) and slowed long before guessing could matter, without ever locking a real link out.

Hooks run one at a time, in a worker thread:

• on_signed_in(link, session) -> state: check the session and store it for link.key. Return "connected", or a state of your own the page has a message for ("wrong_account", say). Anything else, or an exception, is "failed".

• on_ended(link, window_state) -> state: the window ended without a hand-over. By default its state is kept when the page has a message for it (a kept session's "connected"), else "failed".

• before_start(link) -> state | None: runs before each window opens. Return None to go on, or a state of your own to refuse (an invite code the person has not entered yet, say); it may add to link.params.

link.show(name=...) adds public fields to the link's state, which the page's messages can use ({name}).

PageText holds everything the page says ({service} and {label} work in all of it): service, title, help_html, intro_html, button, secure_host (a host it matches gets a padlock), starting, waiting, saving, outcomes (state -> message; with_outcomes(...) adds your own) and script_extra (it may define window.signinBeforeStart, an async function that must return true before a window opens, for an invite code the person enters first, say).

The relay talks to the login service through LoginClient; any object with the same methods (Backend) works, for a login service of another shape.

Security

• The login service never faces the internet: put it on a private network with your server. Every route but /health, noVNC's files and the person's own view needs its bearer (32+ characters); a window's live view needs that window's view token, which only the caller that started it gets, and it is open only while the sign-in page is shown (waiting). The VNC server listens on loopback without a password; the service's WebSocket bridge is the only way in.
• A link's id is its credential. Pages are no-store, no-referrer, not framable, with a strict Content-Security-Policy; nothing in the link's state or page is ever the session.
• Keep the window on the site: signin_window.hosts checks every request against the site's hosts and makes every other host unresolvable for the browser, so neither a link on the page nor a redirect turns the window into a general browser. Lock the browser itself down too ("Locking the browser down"): no address bar, no Chromium pages, no local files.
• A hand-over session leaves the login service once, to your server's hook, and is dropped if nobody takes it.
• Passkeys do not work in a remote window (they need the person's own device nearby): people sign in with a password and a code, and Chromium never offers to keep the password (POLICY).

Report vulnerabilities privately (SECURITY.md).

Development

make bootstrap   # uv sync --locked
make check       # lint, format, tests
uv sync --group browser && uv run playwright install chromium && uv run pytest tests/test_paste_dom.py
NOVNC_DIR=/opt/novnc uv run --group browser python scripts/e2e.py   # with Xvfb and x11vnc installed
sudo .venv/bin/python -m signin_window.browser install-policy && uv run --group browser python scripts/browser_check.py

scripts/e2e.py runs the whole path with real browsers and a stand-in site: a link, its live view, paste, a hand-over and closing the page; CI runs it on every change.

License

MIT

Shared authentication mechanics

The optional signin_window.automation.google.GoogleIdentityDriver supplies a bounded Google sign-in flow through injected browser and secret-filler contracts. Configure its exact controls, observations, allowed origins and callback check; provide a narrowly scoped credential backend explicitly. It reports actions it actually observed. A completed identity callback never proves application access.

automation.secret_filler.OpSecretFiller is an optional 1Password adapter. It checks the top-level origin before and after resolving each field and binds application to an exact frame/document. automation.secrets.OpSecretReader uses only the identity supplied by its caller; a failed read never obtains a different identity. No human window API imports these modules or reads credentials.

For a maintenance process that owns a persistent Chromium profile, install signin-window[chromium] and run patchright install --with-deps chromium. automation.runtime.virtual_display() allocates an isolated Xvfb display. While holding the application's durable profile lease, use automation.runtime.open_browser(profile, display=display) for the browser's lifetime. Its facade supplies origin-bound fills, observations and native X11 input. It preserves the profile, refuses foreign owners and uses a pidfd to bind shutdown to the exact browser it launched. An application still owns network allowlists, identity checks, holds, budgets, publication and scheduling.

Human runners that create their own context can use signin_window.browser.browser_environment(window.env()) to pass only browser runtime settings to Chromium. Operator tokens, planner credentials and ambient proxy variables are excluded. An authorized maintenance runner can also reuse automation.runtime.close_owned_browser after capturing its own process; it never guesses ownership or deletes a profile.