GitHub - aadya940/orbit: Playwright but for any app on any OS

Automate any app, desktop or browser, with composable, typed steps.

Natural language controls the screen. Python controls the flow.

from pydantic import BaseModel
import orbit

class Stories(BaseModel):
    titles: list[str]

async with orbit.session(llm="gemini/gemini-3.6-flash") as s:
    await s.navigate("https://news.ycombinator.com")

    stories = await s.read("the top 5 story titles", schema=Stories)

    if await s.check("a login link is visible"):
        await s.do("click the login link")

Structured extraction into Pydantic models. Real conditionals. Real control flow. No prompt soup.

Why

Agents loop, click the wrong thing, and can't be steered. Orbit splits execution into independent steps (do · read · check · navigate · fill), each with its own model, its own step budget, and typed output.

Goes where browser agents can't. One workflow can read a website, drive a native desktop app, and come back.
Verifies every action. If a click didn't actually change anything, Orbit notices and tries another way instead of confidently marching on.
Sees any app. Cooperative apps are read through the accessibility tree and the live DOM. Canvas apps, custom-drawn UI and remote desktops fall back to pixels automatically.
Cheap model for clicks, powerful model for reasoning. Set llm= per step.
Hard step caps. Nothing runs forever.

Beyond the screen

Orbit ships with tools for the work surrounding a UI task, so one run does the whole job:

async with orbit.session(llm="gemini/gemini-3.6-flash") as s:
    await s.navigate("https://news.ycombinator.com")
    await s.do("read the top 5 stories and save them to stories.csv")

Files (read_file, write_file, read_csv, write_csv, find_files), code (run_python, run_command), clipboard, and web (web_search, fetch_url) are available to the agent by default.

Bring your own tools

from orbit.tools import tool

@tool("post_to_slack", "Send a message to Slack", {"text": {"type": "string"}})
async def post_to_slack(text: str) -> str:
    ...
    return "sent"

async with orbit.session(llm="gemini/gemini-3.6-flash", tools=[post_to_slack]) as s:
    await s.do("summarize this dashboard and post it to Slack")

Any browser

async with orbit.session(llm="gemini/gemini-3.6-flash", browser="firefox") as s:
    ...

chrome (default), edge, brave, opera, vivaldi, chromium, firefox, webkit. Chromium-based browsers use whatever you already have installed; Firefox and WebKit need pip install playwright && playwright install firefox.

Simplest possible use

async with orbit.session(llm="gemini/gemini-3.6-flash") as s:
    await s.do("open the calculator and compute 7 times 8")

Any model works via LiteLLM: set GEMINI_API_KEY, OPENAI_API_KEY, or ANTHROPIC_API_KEY.

Verbs

verb	what it does
`s.do(task)`	perform an action, stop when done
`s.read(task, schema=Model)`	extract typed data, returning a validated Pydantic instance
`s.check(condition)`	returns a real `bool` for `if` statements
`s.navigate(target)`	open a URL or launch an app
`s.fill(form, data)`	fill a form from a dict

Every call takes llm=, max_steps=, guidance= and timeout= overrides.

Results are typed: result.ok, result.output, result.status, result.steps_used, and result.journal, a full record of what happened.

Learn more

example.py: every feature in one file, runnable
examples/: full end-to-end workflows

Windows, Linux and macOS · Python 3.10+

Respect the terms of service of whatever you automate.

Apache 2.0 · Thanks to OculOS