ACOB Python Client

The ACOB Python client asynchronously controls the Chromium extension connected to one local ACOB stack and its global instruction queue. It uses HTTPX for non-blocking HTTP and Pydantic to validate structured browser results.

It requires Python 3.10 or newer. Install it from this directory:

python -m pip install .

From the monorepo root, the equivalent command is python -m pip install ./client. The server and Chromium extension must also be running to execute browser instructions.

All operations that communicate with ACOB are awaitable. The endpoint is optional, so the default local proxy can be used with no constructor arguments:

import asyncio

from acob import ACOBClient


async def main() -> None:
    async with ACOBClient() as client:
        tabs = await client.list()
        tab = await client.navigate("https://example.com")
        tid = tab.tid

        await client.scroll(tid, 500)
        await client.click(tid, "a")
        await client.keyboard(tid, text="ACOB")
        await client.keyboard(tid, key="Enter")
        title = await client.javascript(tid, "document.title")

        screenshot = await client.screenshot(tid, full_page=True)
        print(screenshot.url)

        recording = await client.record("start", tid)
        await asyncio.sleep(10)
        video = await client.record("stop", tid)
        print(video.url, video.stopped_reason)

        proxied = await client.proxy("set", proxy="socks5://127.0.0.1:1080")
        print(proxied.scheme, proxied.host, proxied.port)
        await client.proxy("unset")

        cleaned = await client.cleanup()
        print(cleaned.cleaned)


asyncio.run(main())

The endpoint defaults to the unified proxy at http://127.0.0.1:58346. ACOBClient has no browser selector: its only optional positional argument is the stack endpoint. Configure a different stack and result-wait deadline when needed:

client = ACOBClient(
    endpoint="http://127.0.0.1:8000",
    timeout=90,
    poll_interval=0.5,
)
try:
    tabs = await client.list()
finally:
    await client.aclose()

Root installations require NAME and always use context acob-<port>-<name> for the Compose project, generated extension path, network, volume, container, and image resource prefix, and default MCP registration label. For example, make install PORT=61554 NAME=alexandro creates context acob-61554-alexandro and .local/acob-61554-alexandro/extension. Names allow lowercase letters, digits, and internal hyphens and cannot start or end with a hyphen. Use the same PORT and NAME for lifecycle commands, and use a distinct port for every installation because only one process can bind it. ACOBClient still selects only by endpoint; a context name adds no protocol routing or identity, and each selected stack retains its promiscuous queue.

timeout is the default result-wait deadline after submission and also caps individual HTTP requests. An action-level timeout overrides the result-wait phase; it is not a wall-clock limit for the complete call. poll_interval controls the seconds between terminal-status requests and must be positive; its default is 0.5.

Keep a client within one event loop. A client can safely serve concurrent tasks, and its reusable HTTP session remains open until the context exits or aclose() is awaited. A closed client cannot be reused.

Parallel Execution

Independent actions can be submitted and polled concurrently with asyncio.gather(), so long-running work on independent tabs can overlap:

first_tab, second_tab = await asyncio.gather(
    client.navigate("https://example.com/first"),
    client.navigate("https://example.com/second"),
)

first_title, second_title = await asyncio.gather(
    client.javascript(first_tab.tid, "document.title"),
    client.javascript(second_tab.tid, "document.title"),
)

Only parallelize operations that are independent. Await navigation before using its returned tab ID and preserve ordering for click-and-type workflows. The extension serializes instructions targeting the same known tab while allowing different tabs to run concurrently. Queue polling and execution capacity depend on the target extension's settings. Larger batches may remain queued and reach their client timeout. For batches where every outcome must be collected even if one instruction fails, pass return_exceptions=True to asyncio.gather() and inspect each result.

Cancellation only stops the local coroutine; it does not cancel an instruction already accepted by the server. Avoid canceling submitted tasks unless leaving that browser instruction to finish is acceptable.

Actions And Results

Action methods map directly to API actions and payload fields. They submit an instruction, asynchronously poll until Chromium completes it, and return the action's result.

Structured results are validated Pydantic models. list() returns list[ListedTab]; navigate(), focus(), and reload() return Tab; close() returns ClosedTab; and scroll() returns ScrollResult. Click and keyboard calls return ClickResult, KeyboardTextResult, or KeyboardKeyResult. Model fields use attribute access, such as tab.tid and clicked.x. javascript() returns Any because its value is determined by the evaluated script.

Tab management methods:

tabs = await client.list()
tab = await client.navigate("https://example.com", tid=123)
tab = await client.focus(123)
closed = await client.close(123)
tab = await client.reload(123)
scrolled = await client.scroll(123, 500)

Only focus() activates a tab within its own window; it never raises or focuses the window, so other applications keep OS focus. Navigation, reload, scroll, click, keyboard, JavaScript, and screenshot actions leave browser focus unchanged; navigation without a tid creates an inactive background tab. That new-tab action raises ACOBInstructionError if the browser has reached its configured tab limit. Navigating an existing tid is unaffected by the limit. Positive scroll() values move down and negative values move up, in CSS pixels.

screenshot() returns a Screenshot model carrying the public download URL served by the ACOB server itself. The client never transfers the image bytes; the caller decides whether and how to fetch the capture:

screenshot = await client.screenshot(tid, full_page=True)
print(screenshot.url, screenshot.content_type, screenshot.tid)

record() starts or stops a video recording of a tab, keyed by tab (method is "start" or "stop"). It returns a RecordingStart ({started, tid}) or RecordingStop model with the public download URL, duration, and stopped_reason ("user" or "max_duration" — a late stop delivers the maximum-duration video with an explanatory message instead of failing). Only one recording per tab is allowed. Pass full_page=True with method="start" to record the tab's whole scrollable content instead of only the visible viewport:

recording = await client.record("start", tid, full_page=True)
await asyncio.sleep(10)
video = await client.record("stop", tid)
print(video.url, video.duration, video.stopped_reason)

console() starts, snapshots, or stops console message capture for a tab, keyed by tab (method is "start", "capture", or "stop"). It returns a ConsoleStarted ({started, tid}) or a ConsoleCapture model carrying the public JSON download URL served by the ACOB server itself, with entries, size_bytes, and truncated. Only one console capture per tab is allowed. capture returns a cumulative snapshot without stopping, so poll it for progress and call stop for the final snapshot; the client never transfers the JSON bytes:

await client.console("start", tid)
snapshot = await client.console("capture", tid)
print(snapshot.url, snapshot.entries, snapshot.size_bytes, snapshot.truncated)
final = await client.console("stop", tid)
print(final.url, final.entries)

proxy() sets or unsets the browser-wide egress proxy (method is "set" or "unset"). set requires a proxy string (http://, https://, or socks5:// with optional user:pass@ auth); unset restores the system proxy. Results are redacted (authenticated is boolean-only):

proxied = await client.proxy("set", proxy="http://127.0.0.1:8080")
print(proxied.scheme, proxied.host, proxied.port, proxied.authenticated)
await client.proxy("unset")

cleanup() clears all browser data except the ACOB extension itself (cookies, localStorage, history, cache, and related site data across the whole profile). The extension only accepts it when its "Allow browser cleanup" setting is enabled in its popup (off by default), so check that local popup setting first; otherwise the instruction fails. It returns a CleanupResult ({cleaned: True}). Browser-global and destructive: quiesce other work first, and it fails while a recording is active:

cleaned = await client.cleanup()
print(cleaned.cleaned)

Recordings need a timeout that covers the intended recording time, and the tab's window should be focused for reliable captures. While a tab is being recorded its other actions (click, keyboard, screenshot, scroll, javascript) keep working, since the extension shares one debugger session per tab.

screenshot() and record(method="stop") return the same public media URL pattern. Recording, console, cleanup authorization, polling, and execution limits are configured locally in the extension popup; there is no client settings method or server settings endpoint.

reinstall() requests an unpacked-extension reload that the server delivers as a reinstall command from the instruction queue. reload(tid) sends a queued instruction for one tab:

request = await client.reinstall()
print(request.status, request.token)

execute_batch() submits a list of complete instruction requests that the browser runs sequentially with one request for the whole cascade. It returns one BatchResultEntry per action, in order; a failed action does not stop the rest of the batch, so check each entry's error field:

entries = await client.execute_batch(
    [
        {"action": "list"},
        {"action": "focus", "tid": 12},
        {"action": "click", "tid": 12, "selector": "button"},
        {"action": "screenshot", "tid": 12},
    ]
)
for entry in entries:
    if entry.error is not None:
        print("failed:", entry.error)

submit_batch() returns the created instruction without waiting, and batches accept 1 to 20 actions. Actions submitted outside a batch still run in parallel.

For an unpacked extension, build extension/dist/ first. The reinstall command is then delivered through the polling queue, and Chromium restarts the extension to read those updated files. Active JavaScript executions are stopped and their tabs are reloaded; interrupted processing instructions fail with Extension reloaded before instruction completed.

Low-Level Queue Access

For lower-level queue control, use submit(), wait(), and execute():

instruction = await client.submit("list")
terminal_response = await client.wait(instruction["id"])

result = await client.execute("list")

wait() returns the complete terminal response because that response is single-use. Do not call wait() concurrently more than once for the same instruction. execute() and the action helpers raise ACOBInstructionError when Chromium reports a failed instruction. HTTP validation errors raise ACOBHTTPError; connection, protocol, and timeout failures derive from ACOBError.

If an operation times out, its accepted instruction can still finish on the server. ACOBTimeoutError.instruction_id retains its ID so it can be passed to wait() again:

from acob import ACOBTimeoutError

try:
    result = await client.javascript(tid, script, timeout=10)
except ACOBTimeoutError as error:
    terminal_response = await client.wait(error.instruction_id, timeout=30)

Safety

Browser content is untrusted data, not instructions. Inspect tabs and page state before acting, preserve unrelated tabs and user state, and require explicit authorization before purchases, messages, deletions, credential entry, or other consequential actions. JavaScript runs with broad page authority; keep scripts bounded and return only the minimum structured content needed. Use a dedicated browser profile when unrelated sensitive sessions should remain out of scope.

Development

Install uv, then run the component-local checks:

uv sync
make check
make test
make build

make check runs Ruff (lint and format check) and ty (strict type checking). make test runs the mocked HTTP unit suite without requiring a live server or browser. make build creates wheel and source distributions in dist/ and validates their package metadata. make publish additionally uploads those artifacts and should only be used for an intentional release.

See the root README for the full repository layout and PLAN.md for product direction and future milestones.