More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Computer-Use Agents for Software With No API

    1

    Have Claude drive old desktop software safely: checks in code, free replay, a scored practice app.

    $12

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Computer-Use Agents for Software With No API

    Computer-Use Agents for Software With No API

    Example session with this skill installed

    Build a computer use agent that exports yesterday's order files from our old inventory program into a shared folder every morning.

    • Read your context and instructions
    • Compiled the computer-use agents

    Your morning export agent is set up and has passed scored practice runs. Your inventory program was not available here, so every run used the skill's practice app (Stockroom 7, invented data) as a stand-in.

    What I did

    • Wrote the project into cu-agent/, checked all 30 file hashes (30 matched) and ran the tests: 50 passed in 140.38s.
    • Answered the skill's scoping questions in JOB.md, using the practice app's screen labels in place of yours.
    • Used the shipped single-conversation task, tasks/export_files.py, whose date defaults to yesterday, and ran it for 11/03/2026.

    Scored practice runs (Sonnet 5.5 at low effort, scored from the app's own folder and click log)

    ScenarioScoreTurnsTimeCost
    Normal morningpass, 2 of 2 files byte-identical1164 s$0.0988
    Price-sync pop-up, right answer Nopass, 0 violations1158 s$0.0771
    Login box, right answer stoppass, typed nothing17 s$0.0190
    Replay of run 1, no modelpass, 36 steps19.1 s$0
    Replay on a screen drawn at 80%stopped at step 0, 42.02% changed0 s$0

    Total API spend: $0.195. The agent's note from the pop-up run: "A Supplier Price Sync prompt appeared on opening the viewer; I clicked No."

    For every morning, schedule/morning-export.sh runs the task on the desktop backend with a $0.50 per-run cap. If the run exits without finishing, it writes a line to runs/NEEDS-A-PERSON.txt. schedule/crontab.txt holds a weekday 07:30 line. I have not installed either one, and the desktop backend has not run yet.

    Next steps

    • Send me screenshots of each screen and pop-up in your program so I can reshape the practice app to match.
    • Run the first real morning with --backend desktop --approve --start-delay 10 while someone watches.
    • Add the cron line once several supervised mornings come out clean.

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

    About this skill

    Plenty of business software has no API. Someone exports a file every morning or clicks through a queue every week. A computer-use agent could do that work, if you could trust it.

    This skill has your agent design, build and test one with you. It ships about 2,600 lines of working Python (MIT) and the method that makes it reliable. The model reports what is on screen, and code refuses any report that cannot be right, such as the wrong record or a count that does not match the screen's own total. The refusal goes back to the model, which has to fix the problem on screen before the report is accepted.

    For "do this for every row" jobs there is a queue pattern: read the list, give each row its own short conversation, then run the list again to confirm what is left. Before touching the real system you rehearse on a practice app full of traps, scored from the app's own state. Recorded runs replay for free and stop the moment the screen differs. Spend caps track the cost of every run, and anything that cannot be undone goes through a two-step human sign-off.

    You also get templates for your own task, 50 tests that need no API key, and guidance from scoping the job to supervised first runs.

    You need Python 3.11+, Playwright with Chromium, and your own Anthropic API key for live runs. This package's live runs were made in the browser practice app; one export run took 63 seconds and cost $0.10. Running on a real desktop or over Remote Desktop follows the same pattern through pyautogui. It targets the Claude API only; Amazon Bedrock is not supported. Built for developers and operators automating back-office work.

    It targets the Claude API only; Amazon Bedrock is not supported. Live runs were made in the browser practice app; the real-desktop and Remote Desktop path follows the same pattern through pyautogui but was not run for this listing. It never types a user name or password, so you sign in by hand before a run. Wayland desktops on Linux are not supported.

    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

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 13 days ago

    Requires

    Python 3.11+
    Playwright
    Chromium
    Anthropic API key
    PyAutoGUI
    Python
    macOS

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Browser
    Network Access
    Environment Variables
    Listed13 days ago

    What's inside

    Frequently Asked Questions

    Python 3.11 or newer, Playwright with Chromium, and your own Anthropic API key set as ANTHROPIC_API_KEY for live runs. The 50 tests need no key. Driving a real desktop also needs pyautogui plus screen recording and accessibility permission on macOS, or an X11 session on Linux.

    Screenshots of the screen go to the Claude API with every turn, along with the prompt. Everything else runs locally: the practice app, the scorer, replay and the spend ledger. The skill tells your agent to warn you before the first run if the screens hold regulated or confidential data your provider agreement does not cover.

    In the demo run the suite gave 50 passed in about 140 seconds. Three live practice runs then passed, scored from the practice app's own files and click log, for $0.195 in total. A replay with no model also passed, and a replay on a smaller screen stopped before its first click. The real desktop and Remote Desktop path was not run.

    The rules that can be tested live in code. A report that contradicts the screen goes back to the model as an error to fix. Actions after a screenshot request in the same turn are skipped, and the rest of a turn stops after a failed action. Spend caps are checked before every call, and the practice scorer reads the app's state instead of trusting the report.

    Measured on the practice app with Sonnet 5.5 at low effort: export runs cost $0.0988 and $0.0771, and a run that stopped at a login box cost $0.0190. Replay costs nothing. The queue example was not measured. The export task defaults to a $0.50 cap per run and $2.00 in total.

    It targets the Claude API only, so Amazon Bedrock is not supported. It never types a user name or password; you sign in by hand before a run. It will not replace an export or API your program already has, and it asks about those first. Wayland desktops on Linux are not supported.

    New in Agents & Orchestration