OSWright — desktop automation

    1

    Drive real desktop apps from your agent without paying for a screenshot every step.

    Free

    2 usesSecurity scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    OSWright — desktop automation

    Example session with this skill installed

    Open Calculator and work out 1234 × 56.

    • Read your context and instructions
    • Compiled the oswright

    Calculator is open and shows 69,104.

    I focused the Calculator window, clicked 1, 2, 3, 4, ×, 5, 6 and = with click_element, and read the result back from the app's accessibility tree with get_ui_tree rather than trusting OCR. Between clicks I used observe, which reports only what changed on screen instead of returning a full screenshot each step.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    OSWright is an open-source MCP server for desktop automation — click, type, read on-screen text, manage windows, fill forms. This skill tells your agent when to use it and which tools are the cheap ones.

    The difference from screenshot-based tools: OSWright keeps a model of the screen and re-reads only what changed. Measured on a live desktop, the median step changes 0.012% of the screen's pixels, so a full screenshot is mostly wasted tokens.

    Measured, and reproducible from the repo's benchmarks/:

    • Eight-field data-entry task: 25,122 → 3,404 tokens (7×), with the result checked against the application's own state, not against OCR.
    • Against Windows-MCP on a shared task set: 16.9× cheaper. Windows-MCP was more reliable (20/20 vs 19/20).

    Requires the server: uvx oswright (or pip install oswright). Strongest on Windows, where it uses built-in OCR and UI Automation with no PyTorch download. Linux (X11) and macOS are supported through EasyOCR.

    MIT licensed. Source: https://github.com/Ask-812/oswright

    Strongest on Windows. Linux (X11) and macOS work through EasyOCR, which needs a ~2.5 GB PyTorch download. Electron apps and web content expose little through accessibility, so lookups there fall back to OCR. Needs a live desktop session. In a head-to-head benchmark Windows-MCP was slightly more reliable (20/20 vs 19/20); OSWright was 16.9× cheaper in tokens.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    2 uses

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    • Free with an Agensi account

    Requires

    uv (Python tool runner)
    Windows
    Python
    macOS

    Permissions

    Network Access
    Listed1 month ago

    What's inside

    Frequently Asked Questions

    It keeps a model of the screen between actions and re-reads only the regions that changed, instead of sending a full screenshot after every step. On an eight-field data-entry task this measured 25,122 → 3,404 tokens (7×). Wall-clock time is about the same; the saving is in tokens.

    The skill is fully compatible with Windows via native UI Automation, while macOS and Linux support is provided through an integrated EasyOCR layer for visual text recognition.

    The skill is free. It needs the OSWright MCP server (uvx oswright, or pip install oswright), Python 3.10+, and a live desktop session. It is MIT licensed.

    Often, yes. It works through the screen, the keyboard and Windows UI Automation, so it does not need an API. Apps that draw custom controls without accessibility data are handled through OCR instead.

    Electron apps expose very little through the accessibility tree, so lookups there rely on OCR. This is handled automatically: element lookups fall through a cascade from cached model to OCR to accessibility to a full rescan.

    New in APIs & Backend