Agent Skills
› Lifecycle-Innovations-Limited/claude-ops
› desktop-act
desktop-act
GitHub提供跨操作系统（Linux/macOS/Windows）的桌面自动化能力，支持多Agent并发独占桌面会话、VNC池管理及GUI交互，实现无API密钥的计算机使用控制。
Trigger Scenarios
需要模拟点击或输入以操作本地应用程序
多个Agent需要隔离且并发的桌面环境
Install
npx skills add Lifecycle-Innovations-Limited/claude-ops --skill desktop-act -g -y
SKILL.md
Frontmatter
{
"name": "desktop-act",
"description": "Cross-OS computer-use MCP: multi-agent exclusive desktop leases, Xvnc pool on Linux, single-session macOS\/Windows. Use for GUI\/desktop automation and concurrent agent desktops."
}
desktop-act
Computer-use primitives via mcp__desktop-act__* plus act(goal) (OAuth via
claude-agent-sdk — no Anthropic API key).
Multi-agent (required pattern)
ensure_desktop(owner_id="<agent-id>", geometry="2560x1440") # exclusive seat
# … primitives with session_id …
heartbeat_desktop(session_id) # long runs
release_desktop(session_id) # on stop / idle done
If another agent holds every free seat, Linux auto-spawns a new Xvnc on the
next free display/port. Leases expire (DESKTOP_ACT_LEASE_TTL); reaper frees
the lock and stops pool VNC so idle seats do not thrash the host.
Architecture
- Parent agent is the brain — primitives + screenshots, or hands-off
act(goal). - Linux: X11 + Xvnc/websockify pool (
:50–:99), file-locked leases. - macOS: single interactive session (Screen Sharing → noVNC optional).
- Windows: full WinBackend (ImageGrab + pyautogui/SendInput); single interactive desktop only (no multi-seat pool). Needs a logged-in session.
Watching the desktop
Pool session returns novnc_url (port base DESKTOP_ACT_NOVNC_PORT_BASE,
default 6082). Host default display (often :1) is separate and is never
reaped by the idle reaper. Always prefer the returned URL over hardcoding ports.
Install
- With claude-ops:
desktop-act@ops-marketplace(co-installed by/ops:setupand/ops:updateviaplugin-dependencies.json). - Standalone: marketplace add
Lifecycle-Innovations-Limited/desktop-act, thendesktop-act@desktop-act.
When to use
| Task | Approach |
|---|---|
| Local app you need to click/type into | desktop-act primitives |
| Concurrent agents needing isolated seats | Linux pool + ensure_desktop(owner_id=…) |
| Hands-off goal | mcp__desktop-act__act(goal, max_iterations=…) |
| Browser-only automation | Prefer browser MCP tools when no OS UI is required |
Defaults
actmodel:claude-sonnet-5actmax_iterations:20(notmax_steps)- Python:
>=3.11
Version History
- 64bad13 Current 2026-08-12 09:02
