qirabot 1.4.0

AI automation bolt-on for any framework

Qirabot Python SDK

AI-powered automation SDK that bolts onto your existing browser/mobile automation framework. Let AI see the screen, click, type, extract data, and verify results — with any framework you already use.

Use it two ways: as a Python library (let qirabot launch a Playwright browser for you via bot.open(), or bolt onto a Playwright / Selenium / Appium / Airtest / pyautogui session you already drive), or inside your pytest suite.

Installation

pip install qirabot

Requires Python 3.10+.

The core package has no automation engine of its own — install the extra for the framework you'll drive:

pip install "qirabot[browser]"   # Playwright (needed for bot.open())
pip install "qirabot[desktop]"   # pyautogui (native desktop apps)
pip install "qirabot[appium]"    # Appium (Android / iOS)
pip install "qirabot[airtest]"   # Airtest (Android / iOS / Windows, image-based)

pip install selenium             # Selenium is not an extra — bring your own driver

The airtest extra pulls in Airtest, which pins numpy<2.0 and opencv-contrib-python 4.4–4.6. Installing it into an env that already has numpy>=2 may downgrade or conflict — prefer a dedicated virtualenv.

The Quick Start below uses bot.open(), so it needs qirabot[browser] plus a one-time playwright install chromium. With Selenium you create the driver yourself and pass it to qirabot — see examples/selenium/.

Configuration

export QIRA_API_KEY="qk_your_api_key"

from qirabot import Qirabot

bot = Qirabot()  # reads QIRA_API_KEY from environment

Constructor options:

Parameter	Env Variable	Default	Description
api_key	QIRA_API_KEY	—	API key for authentication
base_url	QIRA_BASE_URL	https://app.qirabot.com	API server URL
timeout	—	120.0	HTTP request timeout (seconds)
model_alias	—	""	Default model alias for all operations
language	—	""	Default response language
task_name	—	""	Optional name for the task (visible in dashboard)
report	—	True	Write an HTML run report (+ screenshots) on close
report_dir	QIRA_REPORT_DIR	""	Output root; default ./qira_runs/<date>/<time-id>/
screenshot_annotate	—	True	Draw a red crosshair at click/type coordinates
screenshot_format	—	"jpeg"	Saved screenshot format ("jpeg" or "png")
screenshot_quality	—	80	JPEG quality, 1–100
retry	—	1	Retries per action on transient failures
retry_delay	—	1.0	Seconds between retries

Model & language

model_alias selects which model backs every operation. The built-in aliases are fast, balanced (the default), and high_quality — trading cost for quality. Check your dashboard for the live list your account can use, then pass the name as model_alias; leave it empty for the default:

bot = Qirabot(model_alias="high_quality")        # applies to all actions
bot.click(page, "Login", model_alias="fast")     # or override per call

language sets the language of AI responses (extracted text, reasoning). It's a short language tag like "zh" or "en" — empty means the server default:

bot = Qirabot(language="zh")                      # extract/ai answers in Chinese
text = bot.extract(page, "获取主标题", language="zh")

Quick Start

This uses bot.open(), so install the browser extra and Chromium first:

pip install "qirabot[browser]"
playwright install chromium

from qirabot import Qirabot

bot = Qirabot()
page = bot.open("https://google.com")

bot.type_text(page, "Search input", "SpaceX", press_enter=True)

summary = bot.extract(page, "Get the first search result title")
print(f"Result: {summary}")

bot.close()

Bind a target (optional)

Every action takes the framework object (page / driver / device / module) as its first argument: bot.click(target, "Login"). When you drive a single, stable target for the whole session, call bot.bind(target) once to get a drop-in proxy that drops the repeated first argument:

bot = Qirabot().bind(driver)     # Selenium/Appium driver, pyautogui, Airtest G/device
bot.click("Login")
bot.type_text("Email", "[e-mail]")
with Qirabot().bind(driver) as bot:   # works as a context manager too
    ...

bind() is recommended for Airtest, pyautogui, Appium, Selenium. For Playwright keep the explicit form page = bot.click(page, ...) so new-tab follows stay visible (a click can open a new tab; the returned page is the one your native page.fill(...) calls should use). With a bound proxy, reach the live page via bot.current_page().

Examples

Runnable examples live in examples/, in two styles:

• Bolt onto your existing tests (pytest) — add AI to a suite you already have: playwright/, selenium/, appium/, desktop/.
• Standalone automation (plain scripts) — scraping / RPA / agents, run with python: automation/.

See examples/README.md for which to pick.

API Reference

Simple Actions

These actions use lightweight vision-based element location — fast and low-cost:

# Click on an element by description
bot.click(page, "Login button")

# Auto-wait: poll until the element looks present (up to timeout) before
# clicking, else raise QirabotTimeoutError. Works on every framework.
# `wait` overrides the auto-derived assertion. (Also on type_text/double_click.)
bot.click(page, "Login button", timeout=15.0, interval=2.0)

# Type text into an input field
bot.type_text(page, "Email input", "[e-mail]")

# Extract data from the screen
text = bot.extract(page, "Get the main heading")

# Verify a visual assertion (returns True/False, never raises)
ok = bot.verify(page, "The success message is visible")

# Wait for a condition (acts as a gate): returns when met, else raises
# QirabotTimeoutError. Use verify() for a non-raising bool check.
bot.wait_for(page, "Page has finished loading", timeout=15.0, interval=2.0)

click, type_text, and double_click return the current target (the same kind you passed in). When an action opens a link in a new tab, the return value is that new tab, so reassign it to keep operating on the active page:

page = bot.click(page, "Open the first video")  # may switch to a new tab

Multi-Step AI (bot.ai())

Let AI autonomously complete a complex task using the full decision engine:

from qirabot import Qirabot, StepResult

bot = Qirabot()
page = bot.open("https://www.google.com")

def on_step(step: StepResult) -> None:
    status = "done" if step.finished else step.action_type
    print(f"  Step {step.step}: {status} {step.params}")

result = bot.ai(
    page,
    "Search for 'best python libraries 2026', click the first result, and extract the main content",
    max_steps=10,
    on_step=on_step,
)

print(f"Success: {result.success}")
print(f"Output: {result.output}")
bot.close()

Screenshot (No AI)

Saves to report_dir/screenshots/ and returns the saved path (or None when report=False):

path = bot.screenshot(page)
print(f"saved to {path}")

Navigation & Scrolling (No AI)

Direct, non-billed actions that don't need AI element location. go_back, navigate, and close_tab return the current page/target (may differ after the navigation); scroll returns None.

bot.navigate(page, "example.com")   # scheme optional; "https://" prepended
bot.go_back(page)                   # back to the previous page (smart, see below)
page = bot.close_tab(page)          # close current tab, return to previous tab
bot.scroll(page, "down", 3)         # scroll at viewport center
bot.scroll(page, "up", distance=5, x=640, y=400)  # scroll at a point

Smart go_back (Playwright): if the current page has back history it goes back in place; if it doesn't — e.g. a click opened a link in a new tab, which starts with no history — and another tab is open, it closes the current tab and returns to the previous one. So the common "click opens a video in a new tab, then go back to the list" loop just works:

for i in range(4):
    page = bot.click(page, locate=f"open video {i + 1}")  # opens a new tab
    bot.screenshot(page)
    page = bot.go_back(page)                               # closes it, back to the list

Reach for close_tab directly when you want to force-close the current tab regardless of history.

Platform support (all actions):

Action	Playwright	Selenium	Appium (mobile)	pyautogui (desktop)	Airtest
click	✅	✅	✅	✅	✅
double_click	✅	✅	✅ ᵃ	✅	✅ ᵃ
right_click	✅	✅	= tap ᵇ	✅	Windows / = tap ᵇ
hover	✅	✅	no-op ᶜ	✅	no-op ᶜ
type_text	✅	✅	✅	✅	✅
clear_text	✅	✅	✅	✅	Android ᵈ
press_key	✅	✅	✅	✅	✅ ᵉ
scroll	✅	✅	✅	✅	✅
drag	✅	✅	✅	✅	✅
navigate	✅	✅	✅	❌	❌
go_back	✅	✅	✅	❌	Android
close_tab	✅	❌	❌	❌	❌
screenshot	✅	✅	✅	✅	✅

AI-located actions (click, type_text, double_click) and the AI operations (extract, verify, wait_for, ai) work on every framework — the matrix shows how each underlying action maps per platform.

• ᵃ Appium/Airtest emulate double_click as two quick taps.
• ᵇ Mobile has no right-click: Appium taps; Airtest right-clicks on Windows only, taps elsewhere.
• ᶜ Touch targets have no hover: Appium/Airtest treat hover as a no-op.
• ᵈ Airtest has no element model; clear_text is best-effort on Android (caret-to-end + repeated delete).
• ᵉ Airtest key names are Android-first (adb); Windows (pywinauto) / iOS use different names.

navigate/go_back raise NotImplementedError where unsupported (pyautogui has no browser-style navigation; Airtest has no URL concept). close_tab is Playwright-only (other targets raise NotImplementedError); the new-tab fallback inside go_back therefore applies to Playwright only — on Selenium/Appium go_back is always history-back, and on Airtest it maps to keyevent("BACK") (Android only; iOS/Windows raise).

Launch a Desktop App (No AI)

pyautogui can drive the mouse and keyboard but cannot open an application. launch_app shells out to the OS so desktop runs can start from a known app:

import pyautogui
from qirabot import Qirabot, launch_app

bot = Qirabot(task_name="wechat")

bot.launch_app("WeChat")              # macOS app name (or bundle id "com.tencent.xinWeChat")
# launch_app("notepad")              # Windows: exe path, registered name, or UWP AppUserModelID
# launch_app("/path/to/app", wait=3) # wait seconds for the window to appear (default 2)

bot.ai(pyautogui, "Send 'hello' to honey in WeChat")

launch_app is also available standalone (from qirabot import launch_app). On macOS it uses open -a/open -b (activating an already-running app), on Windows os.startfile/start/explorer.exe shell:AppsFolder, on Linux the executable directly.

Reports

By default every run writes a self-contained HTML report (with per-step screenshots) when the bot closes — including on error or Ctrl+C, so you can see where it stopped. No model calls, no network; it's built from data captured during the run.

# Default: report on, written to ./qira_runs/<date>/<time-id>/
bot = Qirabot(task_name="checkout")

# Custom output root (date/run subdirs are still added automatically)
bot = Qirabot(report_dir="./artifacts")        # or export QIRA_REPORT_DIR=./artifacts

# Turn it off entirely (nothing written to disk) — e.g. CI / library use
bot = Qirabot(report=False)

Output layout per run:

qira_runs/2026-06-07/192335-3f9ab2c1/
  report.html          # self-contained: embedded thumbnails + PASS/FAIL per ai() task
  screenshots/         # full-resolution frames (click a thumbnail to open)
    001_click.jpg
    002_type_text.jpg
    ...
  recording.mp4        # embedded in the report if you record into report_dir

screenshot_annotate=True (default) draws a red crosshair at the resolved click/type coordinates. To embed a screen recording, point your recorder at bot.report_dir:

dev.start_recording(output=os.path.join(bot.report_dir, "recording.mp4"))

Call bot.report("path.html") to also write the report to a custom location on demand. Use bot.screenshot(target) for a one-off frame (saved under report_dir/screenshots/).

Bolt-On to Any Framework

Qirabot works with your existing automation setup — just pass your page/driver/device object:

Playwright

from playwright.sync_api import sync_playwright
from qirabot import Qirabot

bot = Qirabot()

with sync_playwright() as p:
    browser = p.chromium.launch()
    page = browser.new_page()
    page.goto("https://github.com/trending")

    # Mix playwright selectors with AI
    repos = bot.extract(page, "Get the top 5 trending repo names")
    print(repos)

    browser.close()
bot.close()

Selenium

from selenium import webdriver
from qirabot import Qirabot

driver = webdriver.Chrome()
driver.get("https://www.wikipedia.org")
bot = Qirabot().bind(driver)   # bind once; the driver is stable for the session

summary = bot.extract("Get the first paragraph of the article")
print(summary)

driver.quit()
bot.close()

Android (Appium)

from appium import webdriver
from appium.options.android import UiAutomator2Options
from qirabot import Qirabot

options = UiAutomator2Options()
options.platform_name = "Android"
options.device_name = "emulator-5554"
options.app_package = "com.android.settings"
options.app_activity = ".Settings"
driver = webdriver.Remote("http://localhost:4723", options=options)
bot = Qirabot().bind(driver)

bot.click("Wi-Fi settings")
result = bot.ai("Open Display settings and change font size to Large")
print(f"Success: {result.success}")
bot.close()
driver.quit()

Android / iOS / Windows (Airtest)

Airtest connects to the device itself (no Appium server). G resolves the current device, so bind(G) keeps your usual Airtest style and adds AI on top.

from airtest.core.api import *       # your usual Airtest imports
from qirabot import Qirabot

auto_setup(__file__)                 # your usual Airtest setup, unchanged
bot = Qirabot().bind(G)

bot.click("Login button")            # AI-located — replaces brittle Template images
result = bot.ai("Open Settings and turn on dark mode")
print(f"Success: {result.success}")
touch(Template("native.png"))        # native Airtest still works side by side
bot.close()

Trade-offs and capability notes (e.g. navigate unsupported, go_back Android-only) are in examples/airtest/. You can also pass G, the airtest.core.api module, or an explicit connect_device(...) handle directly without bind().

Error Handling

from qirabot import (
    Qirabot,
    QirabotError,
    AuthenticationError,
    InsufficientBalanceError,
    QirabotTimeoutError,
)

try:
    bot = Qirabot()
    page = bot.open("https://example.com")
    bot.click(page, "Login button")
except AuthenticationError:
    print("Invalid API key.")
except InsufficientBalanceError:
    print("No credits left.")
except QirabotTimeoutError:
    print("Operation timed out.")
except QirabotError as e:
    print(f"Error: {e}")
finally:
    bot.close()

Task Lifecycle

Each Qirabot instance manages a server-side task that tracks all operations:

• Task creation: created when the Qirabot instance is constructed (pass an existing task_id to attach to one instead)
• Step recording: each click(), extract(), ai() call is recorded as a step on the server
• Task completion: call bot.close() or use a context manager — the task is marked as completed
• Auto-cleanup: if close() is not called, atexit ensures cleanup on script exit. The server also has a 30-minute timeout for orphaned SDK tasks.

bot = Qirabot(task_name="my automation")
# ... operations are recorded as steps ...
bot.close()  # task marked as completed

Context Manager

with Qirabot(task_name="my automation") as bot:
    page = bot.open("https://example.com")
    heading = bot.extract(page, "Get the main heading")
    print(heading)
# bot.close() is called automatically

License

MIT