Skip to content

Business operating system for Claude Code — 57 skills, 21 agents, smart daemon. Unified inbox (WhatsApp/Email/Slack/Telegram), autonomous PR merge, full-AWS monitoring, revenue (Stripe+RevenueCat), e-commerce (Shopify), marketing (Klaviyo/Meta/GA4), voice (Bland/ElevenLabs), APM (Datadog/NewRelic/OTEL), YOLO mode.

feat(ops-desktop): /ops:desktop skill + desktop-act MCP wiring

auroracapital24bd57595be4
diff --git a/claude-ops/.claude-plugin/plugin.json b/claude-ops/.claude-plugin/plugin.json
index e4197caa..38e5fcee 100644
--- a/claude-ops/.claude-plugin/plugin.json
+++ b/claude-ops/.claude-plugin/plugin.json
@@ -728,6 +728,34 @@
"title": "Enable FedEx shipping",
"description": "Use FedEx labels via /ops:package. Requires fedex_client_id/secret + account_number.",
"default": false
+ },
+ "desktop_act_command": {
+ "title": "desktop-act runner path (override)",
+ "description": "Absolute path to a custom desktop-act MCP server executable. Leave blank to auto-discover the desktop-act plugin or bootstrap into the per-user cache.",
+ "type": "string",
+ "sensitive": false,
+ "default": ""
+ },
+ "desktop_act_home": {
+ "title": "desktop-act install directory (override)",
+ "description": "Absolute path to a desktop-act marketplace checkout. Used when the plugin is installed outside the standard Claude Code marketplace tree.",
+ "type": "string",
+ "sensitive": false,
+ "default": ""
+ },
+ "desktop_act_repo": {
+ "title": "desktop-act source repository",
+ "description": "Git URL the launcher clones from when bootstrapping a fresh install. Defaults to the public your-org/desktop-act repo.",
+ "type": "string",
+ "sensitive": false,
+ "default": ""
+ },
+ "desktop_act_branch": {
+ "title": "desktop-act branch / tag",
+ "description": "Branch or tag to clone when bootstrapping. Defaults to main.",
+ "type": "string",
+ "sensitive": false,
+ "default": ""
}
}
}
diff --git a/claude-ops/.mcp.json b/claude-ops/.mcp.json
index d665a257..efe682c9 100644
--- a/claude-ops/.mcp.json
+++ b/claude-ops/.mcp.json
@@ -16,6 +16,16 @@
"env": {
"DOPPLER_TOKEN": "${user_config.doppler_token}"
}
+ },
+ "desktop-act": {
+ "command": "python3",
+ "args": ["${CLAUDE_PLUGIN_ROOT}/mcp-servers/desktop-act-launcher.py"],
+ "env": {
+ "DESKTOP_ACT_COMMAND": "${user_config.desktop_act_command}",
+ "DESKTOP_ACT_HOME": "${user_config.desktop_act_home}",
+ "DESKTOP_ACT_REPO": "${user_config.desktop_act_repo}",
+ "DESKTOP_ACT_BRANCH": "${user_config.desktop_act_branch}"
+ }
}
}
}
diff --git a/claude-ops/CHANGELOG.md b/claude-ops/CHANGELOG.md
index 48eec073..bfd6438e 100644
--- a/claude-ops/CHANGELOG.md
+++ b/claude-ops/CHANGELOG.md
@@ -2,6 +2,20 @@
All notable changes to this project will be documented in this file.
+## [Unreleased]
+
+### Added
+
+- **`/ops:desktop` skill + `desktop-act` MCP wiring** — autonomous desktop + browser control surfaced as a first-class ops command. Acquires a noVNC desktop session, takes screenshots, clicks, types, scrolls, and runs the autonomous `act()` loop driven by the bundled `claude-agent-sdk` (OAuth via Claude Max, no Anthropic API key). Cross-platform launcher (Linux full / macOS / Windows partial).
+- **`mcp-servers/desktop-act-launcher.py`** — pure-Python cross-platform launcher. Resolves an installed `desktop-act` companion in (1) `$DESKTOP_ACT_COMMAND`, (2) `$DESKTOP_ACT_HOME`, (3) `$CLAUDE_CONFIG_DIR/plugins/marketplaces/desktop-act`, (4) the per-OS Claude config dir, (5) the per-user cache (`$XDG_CACHE_HOME` on Linux, `~/Library/Caches` on macOS, `%LOCALAPPDATA%` on Windows). If none exist it `git clone`s `$DESKTOP_ACT_REPO` (default your-org/desktop-act), bootstraps a venv, `pip install -r requirements.txt`, then `execv`s the runner. Falls back to a clear error message — never hangs the MCP host.
+- **plugin.json userConfig keys** — `desktop_act_command`, `desktop_act_home`, `desktop_act_repo`, `desktop_act_branch` for per-user overrides.
+- **`.mcp.json` `desktop-act` server registration** — wired via `${CLAUDE_PLUGIN_ROOT}/mcp-servers/desktop-act-launcher.py`. All `mcp__desktop-act__*` tools (`acquire_desktop`, `screenshot`, `click`, `type_text`, `act`, …) are now part of the ops surface.
+
+### Notes
+
+- Linux is the primary supported platform (X11 + Xvnc + websockify + python-xlib). macOS and Windows can launch the server but native X11 calls no-op; browser/Kapture paths still work.
+- First run on a fresh box performs a one-time clone + `pip install`. Subsequent launches `execv` directly into the cached venv with no extra latency.
+
## [2.11.0] — 2026-05-22
First-class Linux parity for the background daemon, plus two cross-platform fixes that surfaced during Linux bring-up on Amazon Linux 2023.
diff --git a/claude-ops/README.md b/claude-ops/README.md
index 04123b05..d17e164a 100644
--- a/claude-ops/README.md
+++ b/claude-ops/README.md
@@ -89,6 +89,7 @@ Per-repo budget caps (default 3/hour), single-flight locks, content-hash dedup,
| `/ops:recap` | **v2** — Status/tail/configure/restart for the multi-session recap marquee daemon |
| `/ops:rotate` | **v2** — Manually rotate the active Claude Max account |
| `/ops:rotate-setup` | **v2** — Multi-account onboarding for the rotator |
+| `/ops:desktop` | Autonomous desktop + browser control via the `desktop-act` MCP companion |
### Dashboard hotkeys
diff --git a/claude-ops/mcp-servers/desktop-act-launcher.py b/claude-ops/mcp-servers/desktop-act-launcher.py
new file mode 100755
index 00000000..1a1f029f
--- /dev/null
+++ b/claude-ops/mcp-servers/desktop-act-launcher.py
@@ -0,0 +1,264 @@
+#!/usr/bin/env python3
+"""desktop-act-launcher — locate or bootstrap the desktop-act MCP server.
+
+Cross-platform launcher (Linux / macOS / Windows) used by the claude-ops plugin
+to bring up the `desktop-act` FastMCP server for the /ops:desktop skill.
+
+Resolution order:
+ 1. ``$DESKTOP_ACT_COMMAND`` — explicit path to an executable script.
+ 2. ``$DESKTOP_ACT_HOME/mcp-server/`` — manual install directory.
+ 3. ``$CLAUDE_CONFIG_DIR/plugins/marketplaces/desktop-act/mcp-server/``
+ 4. Per-OS default Claude config dir.
+ 5. Per-user cache: $XDG_CACHE_HOME/desktop-act-mcp (Linux) /
+ ~/Library/Caches/desktop-act-mcp (macOS) /
+ %LOCALAPPDATA%\\desktop-act-mcp (Windows).
+ If absent and ``DESKTOP_ACT_REPO`` is set, we ``git clone`` it,
+ bootstrap a venv, ``pip install -r requirements.txt``, then exec.
+
+Linux is the only platform with full desktop automation today; macOS and
+Windows currently surface a clear "not supported yet" message so the
+launcher fails loudly instead of hanging the MCP host.
+"""
+
+from __future__ import annotations
+
+import os
+import platform
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+DEFAULT_REPO = os.environ.get(
+ "DESKTOP_ACT_REPO",
+ "https://github.com/your-org/desktop-act.git",
+)
+DEFAULT_BRANCH = os.environ.get("DESKTOP_ACT_BRANCH", "main")
+
+
+def _system() -> str:
+ s = platform.system().lower()
+ if s.startswith("win"):
+ return "windows"
+ if s == "darwin":
+ return "macos"
+ return "linux"
+
+
+def _cache_root() -> Path:
+ sysname = _system()
+ if sysname == "macos":
+ return Path.home() / "Library" / "Caches" / "desktop-act-mcp"
+ if sysname == "windows":
+ base = os.environ.get("LOCALAPPDATA") or str(Path.home() / "AppData" / "Local")
+ return Path(base) / "desktop-act-mcp"
+ base = os.environ.get("XDG_CACHE_HOME") or str(Path.home() / ".cache")
+ return Path(base) / "desktop-act-mcp"
+
+
+def _claude_config_dirs() -> list[Path]:
+ sysname = _system()
+ explicit = os.environ.get("CLAUDE_CONFIG_DIR")
+ out: list[Path] = []
+ if explicit:
+ out.append(Path(explicit))
+ if sysname == "windows":
+ appdata = os.environ.get("APPDATA")
+ if appdata:
+ out.append(Path(appdata) / "claude-code")
+ out.append(Path.home() / ".claude")
+ out.append(Path.home() / ".config" / "claude-code")
+ return out
+
+
+def _runner_in(home: Path) -> Path | None:
+ """Pick the platform-appropriate launcher inside a desktop-act install."""
+ if not home.is_dir():
+ return None
+ if _system() == "windows":
+ for cand in (
+ home / "mcp-server" / "run.cmd",
+ home / "mcp-server" / "run.ps1",
+ ):
+ if cand.exists():
+ return cand
+ sh = home / "mcp-server" / "run.sh"
+ if sh.exists() and os.access(sh, os.X_OK):
+ return sh
+ server_py = home / "mcp-server" / "server.py"
+ if server_py.exists():
+ return server_py
+ return None
+
+
+def _resolve_existing() -> Path | None:
+ override = os.environ.get("DESKTOP_ACT_COMMAND")
+ if override and Path(override).exists():
+ return Path(override)
+
+ explicit_home = os.environ.get("DESKTOP_ACT_HOME")
+ if explicit_home:
+ runner = _runner_in(Path(explicit_home))
+ if runner:
+ return runner
+
+ for cfg in _claude_config_dirs():
+ runner = _runner_in(cfg / "plugins" / "marketplaces" / "desktop-act")
+ if runner:
+ return runner
+
+ runner = _runner_in(_cache_root() / "src")
+ if runner:
+ return runner
+ return None
+
+
+def _exec_runner(runner: Path) -> None:
+ if runner.suffix == ".py":
+ py = (
+ _venv_python(runner.parent.parent)
+ or shutil.which("python3")
+ or shutil.which("python")
+ )
+ if not py:
+ print(
+ "[desktop-act-launcher] No python3 available to exec server.py",
+ file=sys.stderr,
+ )
+ sys.exit(127)
+ os.execv(py, [py, str(runner)])
+ if runner.suffix in (".cmd", ".ps1"):
+ if runner.suffix == ".ps1":
+ ps = shutil.which("pwsh") or shutil.which("powershell")
+ if not ps:
+ print("[desktop-act-launcher] PowerShell not found", file=sys.stderr)
+ sys.exit(127)
+ os.execv(ps, [ps, "-ExecutionPolicy", "Bypass", "-File", str(runner)])
+ os.execv(str(runner), [str(runner)])
+ os.execv(str(runner), [str(runner)])
+
+
+def _venv_python(install_root: Path) -> str | None:
+ bin_dir = "Scripts" if _system() == "windows" else "bin"
+ py = (
+ install_root
+ / ".venv"
+ / bin_dir
+ / ("python.exe" if _system() == "windows" else "python")
+ )
+ return str(py) if py.exists() else None
+
+
+def _have(cmd: str) -> bool:
+ return shutil.which(cmd) is not None
+
+
+def _bootstrap() -> Path | None:
+ """Clone desktop-act into the per-user cache and set up its venv.
+
+ Returns the path to the runner once ready, or ``None`` if bootstrap is not
+ possible (missing git/python3 or unsupported platform).
+ """
+ if not _have("git") or not (_have("python3") or _have("python")):
+ print(
+ "[desktop-act-launcher] git + python3 are required to auto-install desktop-act.\n"
+ " macOS: brew install git python\n"
+ " Linux: sudo apt install -y git python3 python3-venv # or your distro's equivalent\n"
+ " Windows: install Git + Python 3.11+ from python.org",
+ file=sys.stderr,
+ )
+ return None
+
+ cache = _cache_root()
+ src = cache / "src"
+ cache.mkdir(parents=True, exist_ok=True)
+
+ if not src.exists():
+ print(f"[desktop-act-launcher] cloning {DEFAULT_REPO} → {src}", file=sys.stderr)
+ try:
+ subprocess.run(
+ [
+ "git",
+ "clone",
+ "--depth",
+ "1",
+ "--branch",
+ DEFAULT_BRANCH,
+ DEFAULT_REPO,
+ str(src),
+ ],
+ check=True,
+ )
+ except subprocess.CalledProcessError as e:
+ print(
+ f"[desktop-act-launcher] git clone failed (exit {e.returncode}). "
+ f"Set DESKTOP_ACT_REPO to a reachable URL, or install desktop-act manually.",
+ file=sys.stderr,
+ )
+ return None
+
+ py_bin = shutil.which("python3") or shutil.which("python")
+ venv_dir = src / ".venv"
+ if not venv_dir.exists():
+ print(f"[desktop-act-launcher] creating venv at {venv_dir}", file=sys.stderr)
+ try:
+ subprocess.run([py_bin, "-m", "venv", str(venv_dir)], check=True)
+ except subprocess.CalledProcessError as e:
+ print(
+ f"[desktop-act-launcher] venv create failed (exit {e.returncode})",
+ file=sys.stderr,
+ )
+ return None
+
+ vpy = _venv_python(src)
+ if not vpy:
+ return None
+ req = src / "requirements.txt"
+ if req.exists():
+ print(
+ "[desktop-act-launcher] pip install -r requirements.txt (one-time)",
+ file=sys.stderr,
+ )
+ try:
+ subprocess.run(
+ [vpy, "-m", "pip", "install", "--quiet", "-r", str(req)], check=True
+ )
+ except subprocess.CalledProcessError as e:
+ print(
+ f"[desktop-act-launcher] pip install failed (exit {e.returncode})",
+ file=sys.stderr,
+ )
+ return None
+
+ return _runner_in(src)
+
+
+def main() -> int:
+ runner = _resolve_existing()
+ if runner is None:
+ runner = _bootstrap()
+
+ if runner is None:
+ sysname = _system()
+ if sysname != "linux":
+ print(
+ f"[desktop-act-launcher] desktop-act has full desktop-automation support on Linux today; "
+ f"{sysname} support is partial. The MCP server may still expose Kapture / chrome-devtools / "
+ f"Playwright tools but native X11 calls will no-op. Set DESKTOP_ACT_COMMAND to a custom "
+ f"build to override.",
+ file=sys.stderr,
+ )
+ print(
+ "[desktop-act-launcher] Could not locate or bootstrap the desktop-act MCP server.\n"
+ " Install manually: /plugin marketplace add <repo-url> && /plugin install desktop-act\n"
+ " Or pin a custom path: export DESKTOP_ACT_COMMAND=/path/to/run.sh",
+ file=sys.stderr,
+ )
+ return 127
+
+ _exec_runner(runner)
+ return 0
+
+
+if __name__ == "__main__":
+ sys.exit(main())
diff --git a/claude-ops/skills/ops-desktop/SKILL.md b/claude-ops/skills/ops-desktop/SKILL.md
new file mode 100644
index 00000000..b40944db
--- /dev/null
+++ b/claude-ops/skills/ops-desktop/SKILL.md
@@ -0,0 +1,202 @@
+---
+name: ops-desktop
+description: Autonomous desktop + browser control via the desktop-act MCP companion. Acquires an isolated noVNC desktop session, takes screenshots, clicks, types, scrolls, and runs the optional autonomous act() loop. First run auto-bootstraps the desktop-act server into a per-user cache.
+argument-hint: "[goal text | 'status' | 'list' | 'release [session_id]']"
+allowed-tools:
+ - Bash
+ - Read
+ - AskUserQuestion
+ - ToolSearch
+ - mcp__desktop-act__acquire_desktop
+ - mcp__desktop-act__release_desktop
+ - mcp__desktop-act__list_desktops
+ - mcp__desktop-act__list_windows
+ - mcp__desktop-act__launch_app
+ - mcp__desktop-act__screenshot
+ - mcp__desktop-act__observe
+ - mcp__desktop-act__click
+ - mcp__desktop-act__keypress
+ - mcp__desktop-act__type_text
+ - mcp__desktop-act__scroll
+ - mcp__desktop-act__act
+ - mcp__desktop-act__act_step
+ - mcp__desktop-act__batch
+ - mcp__desktop-act__status
+effort: medium
+maxTurns: 40
+---
+
+# OPS ► DESKTOP
+
+`/ops:desktop` drives a real GUI desktop (Linux noVNC pool, partial macOS / Windows support) through the `desktop-act` FastMCP server. Use it for tasks that browser-only automation can't reach: native dialogs, OS file pickers, multi-window flows, screen-recording verification.
+
+The MCP tool schemas are **deferred**. Load them with `ToolSearch` on demand:
+
+```
+ToolSearch select:mcp__desktop-act__acquire_desktop,mcp__desktop-act__screenshot,mcp__desktop-act__click,mcp__desktop-act__type_text,mcp__desktop-act__release_desktop
+```
+
+## Runtime Context
+
+Before any route runs, resolve:
+
+1. **Plugin root**: `${CLAUDE_PLUGIN_ROOT}` — used to locate the launcher.
+2. **MCP launcher**: `${CLAUDE_PLUGIN_ROOT}/mcp-servers/desktop-act-launcher.py` — auto-discovers an installed desktop-act marketplace or, if absent, clones from `$DESKTOP_ACT_REPO` (default `https://github.com/your-org/desktop-act.git`) and bootstraps a venv on first run.
+3. **noVNC default**: `http://<box-host>:6081` — primary desktop. `acquire_desktop()` provisions extra sessions on `:6082+`.
+4. **Mobile / SSH mode**: if `$SSH_CONNECTION`, `$SSH_CLIENT`, or `$SSH_TTY` is set, emit compact text-only output per Rule 7 (no banners, no tables, plain lines). The noVNC URL is the one piece the user definitely needs — print it once on a copy-able line.
+
+## Routing table
+
+Parse `$ARGUMENTS` and route. First token decides:
+
+| First arg | Route |
+| ------------------ | ------------------------------------------------------------------------------------- |
+| (empty) / `status` | Print MCP status, default noVNC URL, active sessions, and platform support note |
+| `list` | `mcp__desktop-act__list_desktops` — show pool state |
+| `release <id>` | `mcp__desktop-act__release_desktop(session_id)` — free a session |
+| `release-all` | Release every session this skill has open this turn |
+| anything else | Treat as a **goal** — acquire → observe → drive → verify → release |
+
+Anything else: treat the whole argument string as the goal.
+
+---
+
+## Route — `status` (default)
+
+1. Probe the MCP via `mcp__desktop-act__status` (auto-load schema first).
+2. Print compact summary — plugin connected? venv ready? pool size? default URL?
+3. If the MCP is **not reachable**, the launcher likely needs to bootstrap. Show:
+ - one-line install command (`/plugin marketplace add <repo> && /plugin install desktop-act`)
+ - or the env override (`export DESKTOP_ACT_COMMAND=/path/to/run.sh`)
+ - or the bootstrap repo (`export DESKTOP_ACT_REPO=https://...`)
+4. Mobile/SSH mode: drop banners, output 4–6 plain lines.
+
+## Route — `list`
+
+```
+ToolSearch select:mcp__desktop-act__list_desktops
+```
+Call it. Format as `session_id · display · vnc · last_used`. Linux only renders full info; macOS/Windows fall back to a "limited platform" note.
+
+## Route — `release <id>` / `release-all`
+
+Per Rule 5 (destructive actions): `release_desktop` tears down the X server. Ask before bulk release:
+
+```
+AskUserQuestion: Release session abcd1234?
+ [Release] [Skip]
+```
+
+For `release-all`, batch confirm once.
+
+---
+
+## Route — **goal** (default for free-form $ARGUMENTS)
+
+The autonomous loop. Use this when Sam types `/ops:desktop open the AWS console and screenshot the ECS cluster page`.
+
+### Step 1 — Acquire
+
+```python
+# Load schemas
+ToolSearch select:mcp__desktop-act__acquire_desktop,mcp__desktop-act__screenshot,...
+
+session = mcp__desktop-act__acquire_desktop()
+session_id = session["session_id"]
+display = session["display"] # e.g. ":51"
+novnc_url = session["novnc_url"] # e.g. http://host:6082
+```
+
+Print the noVNC URL on a line by itself so the user can watch live. On mobile, the URL is the most useful thing — surface it FIRST.
+
+### Step 2 — Observe
+
+```python
+shot = mcp__desktop-act__screenshot(session_id=session_id)
+Read(file_path=shot["path"]) # let Claude see the screen
+```
+
+If you need window-level metadata before clicking blindly:
+
+```python
+windows = mcp__desktop-act__list_windows(session_id=session_id)
+```
+
+### Step 3 — Drive
+
+Two modes:
+
+**A. Streaming primitives** — full reasoning visibility, per-action transparency. Preferred for sensitive flows (anything touching credentials, billing pages, destructive UIs).
+
+```python
+mcp__desktop-act__launch_app(session_id, app="firefox")
+mcp__desktop-act__click(session_id, x=640, y=400)
+mcp__desktop-act__type_text(session_id, text="ECS clusters")
+mcp__desktop-act__keypress(session_id, key="Return")
+mcp__desktop-act__scroll(session_id, direction="down", amount=3)
+mcp__desktop-act__screenshot(session_id) # verify
+```
+
+**B. Autonomous loop** — hands the goal to `act()` which runs claude-agent-sdk against the OAuth-bundled CLI (no API key needed; rides Claude Max). Faster but less observable.
+
+```python
+result = mcp__desktop-act__act(
+ session_id=session_id,
+ goal="$ARGUMENTS",
+ max_steps=20,
+)
+```
+
+For multi-step macros, prefer `batch` (single round-trip) or `act_step` (one micro-step at a time with explicit verification).
+
+### Step 4 — Verify
+
+Always `screenshot` and `Read` the path once at the end before claiming the goal is done. If the verification screenshot doesn't show the expected state, iterate or escalate to the user.
+
+### Step 5 — Release
+
+**Always** release in the same turn, even on failure:
+
+```python
+mcp__desktop-act__release_desktop(session_id=session_id)
+```
+
+If the user wants to keep the session for follow-up actions, ask first via `AskUserQuestion`:
+
+```
+Keep session abcd1234 alive for follow-up?
+ [Keep] [Release]
+```
+
+---
+
+## Cross-platform notes
+
+| OS | Status | Notes |
+| ------- | -------- | --------------------------------------------------------------------- |
+| Linux | Full | X11 + Xvnc + websockify + python-xlib. Multi-desktop pool live. |
+| macOS | Partial | Server runs; native X11 calls no-op. Browser tools still work. |
+| Windows | Partial | Server runs via Python launcher. Native automation requires manual. |
+
+On non-Linux platforms, prefer Kapture (`mcp__kapture__*`) or Playwright for browser-only tasks. The launcher prints a clear "limited platform" notice if it can't bring up a full session.
+
+## Safety
+
+- **Outbound comms** (Rule 6): if a goal involves sending email/messages/forms, stage the draft and ask before pressing Send. Per-message approval, never batch.
+- **Destructive UI clicks**: anything that says "Delete", "Terminate", "Force quit" routes through `AskUserQuestion` first per Rule 5.
+- **Credentials**: never type passwords into a goal string — leave it for Sam to fill in the live noVNC viewer.
+- **Session isolation**: each invocation of `/ops:desktop` gets its own `session_id`. Don't pass session IDs between concurrent agents or skills.
+
+## Examples
+
+```
+/ops:desktop # status
+/ops:desktop list # pool state
+/ops:desktop release abcd1234 # free session
+/ops:desktop open the AWS ECS console and screenshot the my-project-api cluster
+/ops:desktop launch Firefox and navigate to https://finops.lifecycleinnovations.limited
+```
+
+## $ARGUMENTS
+
+$ARGUMENTS