import ezos

ezos.open_app("spotify")                 # launches Spotify (Store app!)
ezos.play_on_youtube("lofi hip hop")     # finds + opens the first video
ezos.search_web("weather", browser="chrome")
print(ezos.weather("London"))            # "London: Overcast +12C, 84% humidity"
ezos.speak("All systems online, boss.")

✨ Why ezos?

When you point an AI at a computer, it needs simple, predictable verbs — not 40 lines of subprocess / ctypes / Win32 boilerplate per task. ezos gives it 85 of them, each named like plain English, each returning a structured result so the AI always knows what happened.

• 🎯 It actually works. App launching uses Windows' own Get-StartApps resolver, so Store apps (Spotify, ChatGPT) and desktop apps (Zoom, Discord) really launch — not just the ones on your PATH.
• 📦 Zero dependencies for the core. Launching, browser, media keys, window control, volume, weather, clipboard, speak — all work on a plain Python install.
• 🤖 Built for AI. Ships OpenAI-compatible + Claude tool formats, a config builder, and a ready-made Jarvis agent loop. Works with OpenAI, Gemini, Z.ai, Groq, and any OpenAI-format API.
• 🔎 Real web research with DuckDuckGo + BeautifulSoup — your AI can find things out, not just open a search page.
• 😌 Readable both ways. ezos.open_app() and ezos.openApp() both work.

📦 Install

git clone https://github.com/<your-username>/ezos.git
cd ezos
pip install -r requirements.txt     # optional extras; the core needs nothing

Or just drop ezos.py next to your script and import ezos.

🚀 Quick start

import ezos

# files
ezos.create_file("notes.txt", "hello world")
print(ezos.read_file("notes.txt"))

# apps & web
ezos.open_app("discord")
ezos.search_web("python tips", browser="chrome")

# system
print(ezos.system_info())
print(ezos.weather("Tokyo"))
ezos.volume_up(10)

# talk & notify
ezos.speak("Task complete")
ezos.reminder(600, "Take a break")       # pops + speaks in 10 min

Every action returns a structured result, so you always know what happened:

ezos.open_app("spotify")
# -> {"ok": True,  "message": "Opened 'Spotify'", "resolved": "Spotify"}
ezos.open_app("notarealapp")
# -> {"ok": False, "message": "Could not find an app called 'notarealapp'. ..."}
#    (also prints a visible  WARNING ezos: ...  so failures aren't silent)

🤖 Build your own Jarvis

ezos is the toolbox; your LLM is the brain. It works with any OpenAI-compatible provider — OpenAI, Google Gemini, Z.ai, Groq, Together, OpenRouter, local servers — just point the client's base_url at them.

1. Configure the AI

aiRead() makes one instruction block. Stack as many as you like and feed them to configure(), which also tells the AI which tools it has, with an instruction for each.

import ezos

cfg = ezos.configure(
    ezos.aiRead("Call me 'boss'."),
    ezos.aiRead("Prefer Chrome for browsing."),
    ezos.aiRead("Research before answering factual questions."),
    provider="openai",
)
cfg["system_prompt"]   # -> system message for your LLM
cfg["tools"]           # -> tools in the OpenAI-compatible format

2. Use the ready-made Jarvis loop

from openai import OpenAI
import ezos

# Google Gemini (via the OpenAI-compatible endpoint)
client = OpenAI(
    api_key="YOUR_GEMINI_KEY",
    base_url="https://generativelanguage.googleapis.com/v1beta/openai/",
)

jarvis = ezos.Jarvis(client, model="gemini-2.0-flash", instructions=[
    ezos.aiRead("Call me boss."),
    ezos.aiRead("Prefer Chrome."),
])

jarvis.chat("open spotify and search lofi on youtube")
jarvis.chat("what's the weather in Tokyo and who won the last F1 race?")

3. Or do raw tool-calling

tools = ezos.openai_tools()           # or ezos.anthropic_tools() for Claude
resp = client.chat.completions.create(model=..., tools=tools, messages=[...])
msg = resp.choices[0].message
if msg.tool_calls:
    messages.append(msg)
    messages += ezos.handle_openai_tool_calls(msg.tool_calls)   # ezos runs them

💡 A complete, runnable chatbot is in test.py (Gemini) and jarvis_demo.py (any provider, with an offline keyword fallback that needs no API key).

🔎 Web research (no API key)

ezos.web_search("who won the match today")       # -> [{title, url, snippet}, ...]
ezos.read_webpage("https://en.wikipedia.org/wiki/Mars")   # clean text
info = ezos.research("python 3.13 new features")  # search + read top pages
print(info["summary"])                            # feed straight to your AI

👁️ Give the AI its own cursor (vision agents)

When commands aren't enough, hand a vision model its own eyes, mouse and keyboard. grid_screenshot() takes a screenshot with a labelled x,y coordinate grid drawn on top — the model reads a coordinate straight off the picture and moves its own cursor there.

cur = ezos.give_cursor()      # the AI's own mouse
kb  = ezos.give_keyboard()    # the AI's own keyboard

shot = cur.look()             # screenshot + coordinate grid -> shot["path"]
# ...show shot["path"] to GPT-4o / Gemini / Claude, it replies "click 840,460"...
cur.click(840, 460)           # click the search bar it saw
kb.type("karan aujla"); kb.press("enter")

Or let the built-in VisionAgent run the whole see → think → act loop for you:

from openai import OpenAI
import ezos

agent = ezos.VisionAgent(OpenAI(), model="gpt-4o")
agent.do("open spotify and play karan aujla")
# each step: grid screenshot -> model picks coordinates -> clicks & types -> repeat

More agent superpowers:

ezos.click_text("Sign in")          # OCR-find on-screen text and click it
ezos.read_screen()                  # OCR everything on screen to text
ezos.remember("owner", "boss")      # long-term memory (survives restarts)
ezos.recall("owner")
cmd = ezos.listen()                 # mic -> text ("play karan aujla on spotify")
ezos.api_call("https://api.github.com/repos/python/cpython")   # hit any API
ezos.run_python("import ezos; ezos.open_app('spotify')")       # AI runs its own code
job = ezos.every(60, "battery_status")    # repeat in the background; job.stop()

For faster tool loops, batch a short sequence of known actions into one model tool call. It runs in order and stops at the first failure by default:

result = ezos.run_batch([
    {"command": "open_app", "args": {"app_name": "spotify"}},
    {"command": "wait_for_window", "args": {"app": "spotify", "timeout": 15}},
    {"command": "look_at_screen", "args": {"app": "spotify"}},
])

wait_for_window(app, timeout=10) waits for a launched app to appear and returns its title and process name. give_cursor(image, 0.5, 0.5) makes a cursor half the normal 48px size; values above 10 set exact pixel dimensions, such as give_cursor("cursor.png", 40, 40).

Use the real tool names in your instructions. A text model can use look_at_screen() to read Windows controls; grid_screenshot() only returns an image file path, so use it with a vision model that receives the image. ai_say("reply") sets up a tag for Jarvis's spoken replies; call speak() to say a startup greeting yourself. If Jarvis stops, inspect jarvis.last_tool_trace to see which tools ran and whether they succeeded.

ezos.give_cursor("cursor.png", 0.5, 0.5)
jarvis = ezos.Jarvis(client, model="glm-4.5-flash", speak_replies=True, instructions=[
    ezos.aiRead("You are Jarvis, a helpful assistant."),
    ezos.ai_say("reply"),
    ezos.aiRead("Open apps with ezos.open_app('name')."),
    ezos.aiRead("Inspect an app with ezos.look_at_screen(app='name')."),
    ezos.aiRead("Click named controls with ezos.click_on(target, app='name')."),
    ezos.aiRead("Move with ezos.move_mouse(x, y); type with ezos.type_text(text)."),
])
ezos.speak("Welcome to Jarvis!")

After a reply, jarvis.last_speech_result reports whether audio playback worked. speak_replies=True makes every final answer speak, even when the model omits the optional <reply>...</reply> tag. Calling set_voice() also enables Jarvis speech automatically; the default remains silent unless requested.

Quiet multi-agent work

Install the optional provider client with pip install ezos[ai]. create_agents starts independent Jarvis conversations on background worker threads and returns immediately from submit(). Supply API keys directly or through the matching environment variable. Providers that expose an OpenAI-compatible API can use their own base_url and model. API keys are not printed or saved by ezos.

pool = ezos.create_agents([
    {"name": "planner", "provider": "openai", "api_key": OPENAI_KEY,
     "model": "gpt-4o-mini"},
    {"name": "reviewer", "provider": "gemini", "api_key": GEMINI_KEY,
     "base_url": "https://generativelanguage.googleapis.com/v1beta/openai/",
     "model": "gemini-2.0-flash"},
])
plan = pool.submit("Suggest a plan for this task.", agent="planner")
review = pool.submit("What risks should I consider?", agent="reviewer")
print(plan.result())
print(review.result())
pool.close()

Agents run quietly in the Python process and do not open their own windows. They have no ezos tools by default, so they can draft and reason without controlling the computer. Give an agent a narrow allowlist such as "tools": ["web_search", "read_webpage"] only when it needs those actions. Tools that control apps or the desktop can still affect your session when explicitly enabled. Call pool.status() to inspect jobs without waiting.

Voice input

Install microphone support with pip install ezos[speech]. voice.take() returns the recognized words; pass an agent to transcribe and immediately send the words to its chat loop:

words = ezos.voice.take(language="en-US")
reply = ezos.voice.take(agent=jarvis)  # transcribe, then jarvis.chat(words)

voice.take() uses the existing listen() speech recognizer and accepts timeout, phrase_limit, and language options. Recognition requires a working microphone and an internet connection for Google's speech service.

For Windows UI inspection and cursor rendering, install pip install ezos[agent]. pip install ezos[all] also gets you Pillow (the grid), pyautogui, OCR (pytesseract — plus the Tesseract engine) and voice (SpeechRecognition). The mouse/keyboard still work dependency-free on Windows via a built-in ctypes fallback.

📚 The ezos commands

Open index.html in a browser for a searchable reference with an example for every function. Categories:

See everything at runtime:

ezos.help_me()            # printed overview
ezos.list_commands()      # list of names
ezos.describe_commands()  # name + params + doc (great for feeding an LLM)

🛠️ How it works

• App launching uses Get-StartApps + explorer.exe shell:AppsFolder\<AppID>, which resolves any Start-menu app (Store or desktop), not just PATH entries.
• Browser finds the real chrome.exe / msedge.exe / firefox.exe so the URL opens in the browser you ask for.
• Media keys, window control, volume use ctypes — no third-party deps.
• Web/weather use the standard-library urllib; research adds DDGS + BS4.

⚠️ Safety

• Built and tested on Windows 11. File/web helpers also work on macOS/Linux; app-launch / volume / window / wallpaper helpers are Windows-first.
• Destructive verbs (delete_file, delete_folder, shutdown, empty_recycle_bin) do exactly what they say — have your AI confirm before calling them. whatsapp_message only opens the chat; you press Send.

🤝 Contributing

Contributions are welcome! To add a command:

1. Write a function in ezos.py that returns _ok(...) / _fail(...).
2. Add a camelCase alias in the aliases block.
3. Add an entry to index.html so it shows in the docs.
4. Open a PR. 🎉

Found a bug or want a feature? Open an issue.

📄 License

Released under the MIT License — see LICENSE. Free to use, modify, and share.