lan-computer-use
/lan-computer-use
Steruje Makiem i komputerami Windows w sieci lokalnej przez SSH: drzewo dostępności, kliknięcia, pisanie, zrzuty ekranu.
- Autor
- silvio-l ↗
- Repozytorium
- skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 8
- Skrypty
- 5
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 5 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/agent/macos/agent.swift, scripts/agent/macos/setup.sh, scripts/agent/windows/agent.cs, scripts/agent/windows/setup.ps1, scripts/cu.py
Co robi i kiedy
Steruje Makiem i komputerami Windows w sieci lokalnej przez SSH: drzewo dostępności, kliknięcia, pisanie, zrzuty ekranu.
Wyzwalacze z opisu
Oryginalny opis (EN)
Fast GUI control of this Mac or LAN Macs and Windows PCs over SSH via accessibility trees with element refs, clicks, typing and screenshots. Use for 'Rechner steuern', 'Computer Use', 'klick in der App', 'Windows-PC bedienen'.
Jak działa
- Workflow
- Rules
Podgląd SKILL.md
---
name: lan-computer-use
description: "Fast GUI control of this Mac or LAN Macs and Windows PCs over SSH via accessibility trees with element refs, clicks, typing and screenshots. Use for 'Rechner steuern', 'Computer Use', 'klick in der App', 'Windows-PC bedienen'."
---
# LAN Computer Use
Drive desktop apps on this Mac (`local`) or on Macs and Windows PCs in the LAN. One CLI talks to a small agent that runs in the target's logged-on desktop session. The agent is compiled on first use and reached over an SSH tunnel; a local agent needs no SSH. It reads the accessibility tree (UI Automation / AX) and acts on elements by ref, which is faster and more reliable than working from pixels.
```sh
"${CLAUDE_SKILL_DIR}/scripts/cu" snapshot -i # this Mac
"${CLAUDE_SKILL_DIR}/scripts/cu" -t win snapshot -i # a LAN machine (targets.json entry or SSH host alias)
```
Call `scripts/cu` with its full path in every command. Don't put the command in a shell variable (zsh does not split words, so `$CU snapshot` fails), and don't copy the script elsewhere: the path stays valid across updates. Below, `cu` stands for that full path.
## Workflow
1. **Announce.** Before the first input on a machine, tell the user in text which machine you will control and why. On their own Mac, also say it aloud (`say "…"`); `cu` does this once per session for targets with `"announce": true`. Then take a control lease, which also shows the see-through overlay (session name, task, current action, pressed keys, countdown, and below it your own last one or two chat or tool steps):
`cu task "Export im Dialog prüfen" --eta 3m`
`--eta` is required and **at most 5m**: plan control in short sessions of 5 minutes or less. Estimate realistically (count the steps, about 3–10 s each when you work compactly, plus app starts and waits). Without a valid lease the agent refuses every input; 60 s after the lease ends it quits by itself. When `cu` warns that the lease ends soon and work is left, renew it with `cu task --eta …` (again at most 5m), otherwise finish.
2. **Look.** `cu snapshot -i` prints a flat list of clickable or editable elements of the foreground window, each with a ref:
`[e7] button "Speichern"`. Pick a window with `-w <title>` or `-a <app>`. Add `--rects` only when you need positions.
3. **Act by ref.** Use `click e7`, `type "text" --ref e3`, `set-value e3 "text"`, `press e7`, `key cmd+s`. Every input command answers with the state after the action, i.e. the foreground window and the focused element. Read that line instead of taking a new snapshot.
The overlay steps aside by itself when you move the mouse onto it. Need to see a spot it covers without moving there? Say so in text, then `cu hide 20s` (announced in the panel, max 60 s, not in the first 5 s); `cu hide 0` ends it early.
4. **Verify cheaply.** Use `find "Gespeichert"` or `wait "Fertig" --timeout 20`, then `snapshot -i` again. Take `shot` only when the meaning is visual: layout, images, canvas, or a dialog without an accessibility tree. Read the PNG path it prints.
5. **Release at once.** Right after the last input, run `cu task --done` and then `cu down`. Do this before you analyse results, write a report or ask the user anything, so the overlay and the agent never stay up while you think. `down` stops the agent, removes it from the target and hands the screen back. If you need the machine again later, start over at step 1.
Refs (`eN`) are valid until the next `snapshot`/`find` on that machine. A stale ref gives a clear error; then take a new snapshot. No ref for the target? Use `click --text "Label"` (on macOS it falls back to OCR), then `shot` + `click --img X Y` with pixel coordinates from that shot.
## Rules
- **Fluent and compact.** The person watches a screen that should keep moving. Chain steps with `batch` or back-to-back calls, use the post state of each reply instead of extra snapshots, never `sleep` between steps (use `wait` for real waits), keep `--settle` small and don't pause to think mid-session: plan first, then act in one go.
- Never type secrets. `type`/`set-value` refuse password fields; `--password-ok` is only for the user's explicit consent in this conversation.
- Prefer `press`/`set-value` (no mouse, works in the background) over coordinate clicks; prefer `snapshot -i` over `shot`.
- Chain many steps without a reconnect: `batch` reads JSON lines (`{"cmd":"click","ref":"e3"}`) from stdin.
- `hello` shows the OS, the user, whether the desktop is locked or on a secure desktop (UAC), the screen and missing permissions. Run it first when anything fails.
- Websites: use a browser automation tool, not this skill. This skill is for native apps and OS dialogs.
Read `${CLAUDE_SKILL_DIR}/REFERENCE.md` for target setup (`targets.json`, SSH, permissions), every command and flag, the snapshot format and platform quirks.
Pliki w paczce (8)
- REFERENCE.md 14 KB
- SKILL.md 5 KB
- scripts/cu 1 KB
- scripts/cu.py 50 KB
- scripts/agent/macos/agent.swift 72 KB
- scripts/agent/macos/setup.sh 3 KB
- scripts/agent/windows/agent.cs 75 KB
- scripts/agent/windows/setup.ps1 4 KB
Wersje
W bazie nie ma innych skilli o nazwie lan-computer-use.
Migawka bazy z 13.11.2026 · wersja paczki 48c7c7ef54fb (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100computer-use
Steruje interfejsem graficznym widocznego okna lokalnej aplikacji przez orca computer: drzewo dostępności, kliknięcia, pisanie, menu, dialogi i zrzuty ekranu. - 98windows-pc-control
Obserwuje ekran komputera z Windows i steruje aplikacjami przez kliknięcia, wpisywanie tekstu i klawisze, gdy trzeba wykonać czynność w programie. - 100computer-use
Steruje interfejsem graficznym widocznego okna lokalnej aplikacji przez manta computer: drzewo dostępności, kliknięcia, pisanie, menu, dialogi i zrzuty ekranu. - 100terminal-mac
Steruje odizolowanym wirtualnym macOS przez zrzuty, kliknięcia i skróty, by testować aplikacje bez zmian na pulpicie. - 98computer-use-windows
Obsługuje interfejs aplikacji Windows przez computer_use MCP: obserwuje ekran, klika i pisze na wskazanym komputerze, także zdalnym. - 96computer-use
Steruje natywnymi aplikacjami na komputerze przez CuaDriver, w tym kalkulatorem, drzewem dostępności i zrzutami ekranu.