sweef-browser

Python SDK for Sweef, a browser engine for AI agents written in Go. Synchronous, standard library only, typed, Python 3.8+. The API follows Playwright's sync API where it can, and pages are read as a YAML accessibility tree in which every control carries a [ref=eN] handle you can pass back as a selector.

The SDK talks to the sweef engine over a local socket; it installs the engine for you on first use, so this is all you need:

pip install sweef-browser

The sweef engine: installed for you

You do not download Sweef or edit PATH. The SDK finds the engine in this order:

1. the binary option (Sweef.launch(binary=...))
2. $SWEEF_BIN
3. sweef on PATH
4. the per-user cache: $XDG_CACHE_HOME/sweef/<engine-version>/sweef (default ~/.cache/...), ~/Library/Caches/sweef/... on macOS, %LOCALAPPDATA%\sweef\cache\<engine-version>\sweef.exe on Windows
5. otherwise it downloads the pinned engine (0.1.0-alpha.6) from the GitHub release into that cache, once, and prints one line to stderr. The archive's SHA-256 is checked before anything is executed; the SDK pins the expected hashes for its default version and falls back to the release's SHA256SUMS for other versions.

pip install sweef-browser (or pipx install sweef-browser) also puts a sweef command on PATH (it runs the same engine, downloading it first if needed). sweef-install or python -m sweef downloads the engine ahead of time and prints where it is (--print-path only looks, never downloads). import sweef has no side effects; the download happens on the first Sweef.launch() / Sweef.session(). In code, sweef.ensure_binary() returns the engine path, downloading it if necessary.

Air-gapped: mirror the release files under $SWEEF_DOWNLOAD_BASE/v<version>/, or put the engine on PATH / in SWEEF_BIN and set SWEEF_SKIP_DOWNLOAD=1. Uninstall: delete the cache directory (<cache>/sweef/<engine-version>); nothing else is written outside the package, and PATH is never modified.

Quick start

from sweef import Sweef

with Sweef.launch() as sw:                       # starts `sweef serve` on a private socket, stops it on exit
    page = sw.new_page("https://example.com")
    print(page.title())                          # Example Domain
    print(page.snapshot(interactive_only=True))  # - link "Learn more" [ref=e1]: ...
    png = page.screenshot()                      # PNG bytes
    print(len(png), "bytes of PNG")

Connecting

Pages

import re
from sweef import Sweef

with Sweef.launch() as sw:
    page = sw.new_page("https://news.ycombinator.com/")
    tree = page.snapshot()                                    # options: interactive_only, delta, selector, max_chars
    ref = re.search(r'link "new" \[ref=(e\d+)\]', tree).group(1)
    page.click(ref)                                           # refs, CSS, text=, role=, label=, placeholder=, testid=
    page.get_by_role("link", name="past").click()
    print(page.snapshot(delta=True))                          # only the lines that changed

• Reading: title(), url(), content(), text_content(sel), snapshot(...), accessibility_tree(), screenshot(path=, full_page=, width=, height=), screenshot_text(), pdf(path=, format=), frames(), popups().
• Acting: click, fill, type, press (keys and combinations such as Control+a), check, uncheck, select_option, hover, focus, tap, drag_and_drop, scroll, keyboard.type/press, mouse.click/move/down/up/wheel.
• Locators: page.locator(sel) with click, fill, text_content, inner_html, get_attribute, input_value, count, is_visible, is_checked, bounding_box, wait_for, nth, first, last.
• JavaScript: evaluate(expr) or evaluate("s => s.length", arg), add_init_script, expose_binding + binding_calls.
• Waiting (the page keeps running meanwhile): wait_for_selector, wait_for_function, wait_for_load_state, wait_for_url, wait_for_timeout.
• Network: route(glob, json=... | body=..., status=, action=), unroute, unroute_all, routes, network_log, har, set_extra_http_headers, downloads, download.
• Logs and dialogs: console_messages, events, arm_dialog("accept"), dialogs.
• Emulation: set_viewport_size, set_user_agent, set_geolocation, grant_permissions, set_offline, emulate_media(color_scheme).

Batches: many steps, one round trip

from sweef import Sweef

with Sweef.launch() as sw:
    page = sw.new_page()
    r = page.batch([
        "open https://the-internet.herokuapp.com/login",
        "fill #username tomsmith",
        "fill #password SuperSecretPassword!",
        "click button[type=submit]",
        "expect text 'You logged into a secure area'",
        {"method": "Page.title"},                     # any RPC method works as a step
    ], observe="interactive")
    print(r["ok"], r["totalMs"], r["observation"])

Contexts

ctx = sw.new_context(user_agent="my-agent/1.0", viewport={"width": 390, "height": 844})
p = ctx.new_page("https://example.com")
print(ctx.cookies())
ctx.storage_state(path="state.json")
ctx.close()

Every RPC method of the engine is reachable with sw.call(method, params); see the method list in the main README.

Tests

SWEEF_BIN=/path/to/sweef python -m unittest discover -s tests runs the suite against a real sweef serve and a local HTTP fixture (no network).

License: Apache-2.0.