TELEGRAM_API_ID — Numeric API ID from https://my.telegram.org/apps (create a personal app, not a bot).
TELEGRAM_API_HASH — API hash from the same my.telegram.org/apps page.
TELEGRAM_PHONE — Your Telegram phone number in E.164 format (e.g. +14155551234).
TELEGRAM_SESSION — gram.js StringSession token. Generate once via: node index.js --auth (prompts for SMS code + optional 2FA). Store in your secret manager — never commit.
Claude Code ไม่มีลิงก์ติดตั้งสำหรับคำสั่งแบบโลคัล ให้คัดลอกคำสั่งด้านล่างแล้วรันในเทอร์มินัล
Turn Claude Code into a complete business operating system — infrastructure health, CI/CD status, unified inbox, open PRs, sprint state, revenue snapshot (Stripe + RevenueCat + AWS), and autonomous C-suite agents that act on your behalf.
What's new in v3.6
Current: see the version badge above and the latest release. One plugin, three harnesses, official skill shape.
Hermes
Installer mirrors skills into ~/.hermes/skills for working slash commands and links hermes-plugin/ for skill_view("ops:*").
Grok
Loads the Claude plugin as-is. No .grok-plugin.
Skills
Third-person trigger descriptions, shared preamble, ops-rules (plugin-root CLAUDE.md is a pointer). Oversized skills split into references/.
v2 turns claude-ops from a briefing + comms surface into an autonomy layer for Claude Code itself. Purely additive — no v1 behaviour changes by default. See claude-ops/CHANGELOG.md and docs/migrating-from-v1.md.
Multi-account Claude Max rotator with launchd daemon + AI-brain (jittered post-rotation respawn + /login re-auth so a freshly-rotated account isn't hit by a synchronized first-call burst)
# 1. Upgrade
/plugin update ops@ops-marketplace
# 2. Run the wizard (hits new steps 6.5a–6.5d for v2 toggles)
/ops:setup
# 3. Map your repos to their deploy URLs
/ops:deploy-fix configure
# (opens ~/.claude/config/post-merge-services.json)
# 4. From now on, every `gh pr merge` you run from Claude Code will:
# - poll the deploy workflow
# - curl /health on success
# - verify /version returns the merged SHA
# - on failure: auto-rerun transients, OR dispatch a Haiku deploy-fixer
/ops:deploy-fix # see status / budget / live runs
Per-repo budget caps (default 3/hour), single-flight locks, and content-hash dedup prevent runaway spending. Notifications route via macos/ntfy/pushover/discord/telegram/none. Every toggle is spacebar-toggleable in /plugins settings.
Quick Start
# 1. Add the marketplace (ships ops + desktop-act companion)
/plugin marketplace add Lifecycle-Innovations-Limited/claude-ops
# 2. Install the ops plugin
/plugin install ops@ops-marketplace
# Optional explicit companion (also run by /ops:setup and /ops:update):
# /plugin install desktop-act@ops-marketplace
# 3. Run the guided setup wizard (co-installs desktop-act by default)
/ops:setup
git clone https://github.com/Lifecycle-Innovations-Limited/claude-ops.git
cd claude-ops/installer
npm install
node bin/claude-ops-installer.mjs install
One command mirrors upstream skills + binstubs into every detected CLI's expected layout from a single central config (~/.config/claude-ops-installer/config.yaml). See installer/README.md for the schema, supported agents, and verify / doctor / update / uninstall subcommands.
Already installed? Check that every CLI loads the same skills (read-only):
ops-update --check
# or, from a clone without npm dependencies:
node claude-ops/lib/parity/check.mjs
It compares the bytes Claude Code, Codex, Grok, Cursor and Hermes actually load
against the installed release and writes, fetches and locks nothing. Every
status, exit code and fix is in claude-ops/docs/skill-parity.md.
[!TIP]
The wizard installs the background daemon EARLY (Step 2c). While you're still answering "connect Slack? [OAuth/Skip]" questions, briefing-pre-warm is already running every 2 minutes — pre-fetching ECS health, git state, PRs, CI, and unread counts. By the time setup finishes, your first /ops:go briefing loads in <3 seconds from warm cache instead of ~30s cold.
Local development:
git clone https://github.com/Lifecycle-Innovations-Limited/claude-ops.git
claude --plugin-dir ./claude-ops/claude-ops
/ops:voice is a full voice surface — not just AI calls.
Subcommand
What it does
phone <number>
Native Phone.app dial via Continuity
facetime <handle> [--audio]
FaceTime video or audio
zoom start|join|schedule
Native Zoom + REST scheduling
join [--at now|next]
Auto-joins the current/next calendar meeting with smart AV defaults (cam/mic policy by attendee count, lid-state mic switching, Elgato auto-launch)
twilio-call, twilio-sms
Programmatic outbound voice + SMS
bland-call
AI agent phone call
tts, transcribe
ElevenLabs TTS + Groq Whisper
All outbound 1:1 channels (twilio, bland) go through the per-message approval gate (Rule 6).
Skill routing
flowchart TD
U["User"] --> O["/ops"]
O --> D["Dashboard"]
D --> Daily["Daily Ops"]
D --> Eng["Project & Eng"]
D --> Biz["Business"]
D --> Auto["Automation"]
Daily --> go["/ops:go"]
Daily --> inbox["/ops:inbox"]
Daily --> merge["/ops:merge"]
Eng --> projects["/ops:projects"]
Eng --> linear["/ops:linear"]
Eng --> fires["/ops:fires"]
Biz --> revenue["/ops:revenue"]
Biz --> ecom["/ops:ecom"]
Biz --> marketing["/ops:marketing"]
Auto --> yolo["/ops:yolo"]
Auto --> orchestrate["/ops:orchestrate"]
Before / After
┌────────────────────────────────────────────┬──────────────────────────────────────────────┐
│ WITHOUT claude-ops │ WITH claude-ops │
├────────────────────────────────────────────┼──────────────────────────────────────────────┤
│ Open 6+ tabs every morning │ /ops:go —— one command, done │
│ Context-switch between Slack/Telegram/ │ /ops:inbox —— unified view, all channels │
│ email │ │
│ Manually review and merge PRs one by one │ /ops:merge —— autonomous pipeline │
│ SSH into servers to check health │ /ops:fires —— terminal dashboard │
│ Forget to track AWS spend │ /ops:revenue —— automatic cost snapshot │
│ Switch between Linear and GitHub │ /ops:linear + /ops:projects —— unified │
└────────────────────────────────────────────┴──────────────────────────────────────────────┘
Integrations (22 services)
Most integrations offer two paths — MCP (zero-config OAuth) or CLI (fuller feature set). The setup wizard lets you choose per-integration.
SERVICE
MCP
CLI
WHAT YOU LOSE WITHOUT CLI
GitHub
—
gh (auto)
EVERYTHING — CI logs, PR merge, triage all require gh
AWS
—
aws (auto)
EVERYTHING — 17+ services probed by infra-monitor
Stripe
—
API key
Required for /ops:revenue MRR — web + desktop subs
Virtual camera setup for meetings via /ops:voice join
GSD
—
auto-detected
Optional — roadmap state; degrades gracefully
Doppler
@dopplerhq/mcp-server (MCP)
doppler CLI (fallback)
Secrets manager; MCP server provides direct tool access
[!NOTE]
infra-monitor now covers every AWS service you have IAM for — ECS, EC2, RDS, Lambda, S3, CloudFront, ALB/NLB, API Gateway, SQS, SNS, DynamoDB, ElastiCache, Route 53, ACM, CloudWatch, Budgets, IAM. Probes run in parallel; services you can't access are silently skipped.
Architecture
flowchart TB
CC[Claude Code] --> S[Skills · 22]
CC --> A[Agents · 12]
CC --> H[Hooks]
S & A & H --> RC[Runtime Context]
RC --> P[preferences.json]
RC --> M[memories/]
RC --> SEC[Doppler · PW Mgr · Keychain · env]
CC <--> D[ops-daemon<br/>launchd · 7 services]
All skills use pre-execution shell blocks (! fences) that gather data before model context loads — zero extra latency, minimal token overhead. The ops-daemon pre-warms briefing data so /ops:go hits warm cache.
Why the nested claude-ops/claude-ops/ directory? Claude Code's plugin marketplace system requires a two-level layout: the repo root acts as a marketplace container (with .claude-plugin/marketplace.json pointing "source": "./claude-ops"), while the inner directory is the actual plugin root (with .claude-plugin/plugin.json, skills, agents, etc.). This is how Claude Code resolves and caches plugins — it cannot be flattened.
claude-ops/ ← marketplace root (this repo, this README)
├── .claude-plugin/
│ └── marketplace.json # points to ./claude-ops as plugin source
├── README.md # ← you are here
│
└── claude-ops/ ← plugin root (Claude Code loads from here)
├── .claude-plugin/plugin.json
├── CLAUDE.md # 5 non-negotiable plugin rules
├── skills/ # 22 slash commands
├── agents/ # 12 autonomous agents (Opus/Sonnet/Haiku)
├── bin/ # ops-gather · ops-shopify-create · gog fallback
├── hooks/ # SessionStart health check
├── telegram-server/ # bundled MCP server (gram.js)
├── templates/ # Shopify Admin + app scaffolding
├── tests/ # bash validation · test-no-secrets.sh
└── .mcp.json # empty on purpose — MCP starts on demand, not per session
Agent Teams
Every ops skill that spawns agents supports Claude Code Agent Teams — a coordination layer where agents share context, report progress, and accept mid-flight steering.
Enable: Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in your environment.
How it works: When the flag is set, skills create a named team and dispatch agents into it. Agents within a team can share findings (e.g., an inbox agent discovers a Slack message referencing an email thread, so the email agent prioritizes it) and you can steer priorities via SendMessage.
TeamCreate("fire-fixers")
Agent(team_name="fire-fixers", name="fix-ecs", ...)
Agent(team_name="fire-fixers", name="fix-ci", ...)
SendMessage(to="fix-ecs", content="This is P0, prioritize over CI")
Without the flag: Skills fall back to standard fire-and-forget subagents — still parallel, but no coordination or steering.
Compliance enforced by CI:tests/test-agent-teams.sh audits every skill for Agent Teams support — any skill with Agent in its allowed-tools must have TeamCreate/SendMessage, a documentation section, the feature flag check, and a fallback path.
Competitor Intelligence (v2.3)
A self-discovering competitive-intelligence pipeline that goes well past "weekly Google Alert." Free signals + LLM curation, $0 incremental cost, configurable per brand.
Pipeline per configured brand:
Discovery (Tavily, cached 30d) — surfaces the current competitor landscape for {brand_name} in {category}.
Per-competitor signal collectors run in parallel (scripts/lib/competitor/):
reddit-search.sh — Reddit JSON API
hn-search.sh — HN Algolia API
appstore-lookup.sh — iTunes Lookup (opt-in for mobile brands)
jobs-feed.sh — Greenhouse + Lever public APIs (senior-hire detection)
page-diff.sh — HTML pricing/features/careers SHA snapshots with money-token detection
Severity routing via event-router.sh:
high (price change on direct rival, funding, Show HN going viral) → immediate Telegram push every 10 min via competitor-alert cron
med → daily 17:00 grouped roll-up via competitor-daily cron
low → state-only, surfaces in weekly strategic synthesis
Weekly synthesis (Mon 10:00) — claude_invoke Sonnet 4.6 against 7-day events.jsonl window. Output: NEW entrants / Competitor moves / Brand signal / Threats & opportunities.
Required env:TAVILY_API_KEY (free tier 1000 searches/mo — covers ~30 brands). Optional:TELEGRAM_BOT_TOKEN + TELEGRAM_CHAT_ID for push delivery (otherwise reports persist to disk only).
Cost at 10-brand scale: ~13 Tavily calls/wk + ~320k Sonnet tokens/mo on Max-OAuth = $0 incremental.
Privacy & Security
[!IMPORTANT]
Transparency matters. claude-ops reads from your AWS, GitHub, Linear, Sentry, WhatsApp, Email, Slack, Telegram, Shopify, Stripe, RevenueCat, and more. You should know exactly what it touches.
Credential resolution chain (in order): Doppler MCP → Doppler CLI → 1Password/Dashlane/Bitwarden → macOS Keychain → env vars → Claude Code's encrypted userConfig (~/.claude.json).
Setup auto-scan sources (only during /ops:setup): env, shell profiles, Doppler, 1Password, Dashlane, Bitwarden, macOS Keychain, Claude Code's ~/.claude.json, Chrome history URL list (never page content), Slack Playwright profile (only if chosen).
The plugin does NOT:
Phone home. No telemetry. No analytics. No crash reports.
Upload data to any third party you haven't configured.
Access clipboard, camera, microphone, or SSH keys.
Perform disk-wide scans — every scan is a targeted path.
Background daemon services (only those you enable):
briefing-pre-warm every 2 min — parallel ops-gather for ECS/git/PRs/CI/unread. Local only.
wacli-sync continuous — WhatsApp Web protocol, same as standalone wacli.
memory-extractor every 30 min — Haiku summarizes local chats to memories/.
inbox-digest every 4h — aggregates for your configured Telegram bot (if any).
Just Claude Code 1.0+. Everything else is installed automatically by /ops:setup via Homebrew (macOS), apt (Linux), or winget (Windows). /ops:speedup auto-detects macOS / Linux / WSL / Windows and applies host-appropriate tuning (no manual flags needed).
What's New in v1.7.0
/gtm — cross-channel go-to-market planner (NEW skill). Strategy layer on top of /ops:marketing that generates plans across paid, unpaid, sales, and AI-automation avenues and hands launchable items to /marketing via the Skill tool.
/ops:projects portfolio dashboard — every project in the GSD registry with active phase, task count, dirty-file count, and open-PR status. Backed by the gsd-registry-sync daemon service.
ops-speedup v2 parity — --gpu (Neural Engine + GPU util via powermetrics), --power (energy hogs from top -o pmem), --os-actions (cross-platform kernel_task / WindowServer restarts + launchd/systemd masking behind an allowlist). Hardened against 9 review findings including a SEV-9 eval shell-injection and a SEV-8 RETURN-trap race.
ops-memory-extractor Claude Code OAuth support — prefers the OAuth token stored in the macOS Keychain (Claude Code-credentials) so memory extraction is billed against the Claude Max subscription instead of the API credit. Falls back to ANTHROPIC_API_KEY. The OAuth token is never exported to the shell.
Persistent WhatsApp --follow — wacli-keepalive.sh no longer tears down the follower within 5-20 min of start. INITIAL_BACKFILL_DELAY=30 lets the follower stabilize before the first --once sweep, and a reentrant guard prevents overlapping sweeps.
MCP auto-reconnect — PreToolUse hook kills and respawns any disconnected MCP server without user prompting.
57 skills, 21 agents — up from 21/12 in v0.6.0. Full list in claude-ops/README.md.
Models: C-suite on Opus 4.6, scanners/monitors/fix agents on Sonnet 4.6, memory extractor on Haiku 4.5.
Contributing
PRs welcome — see CONTRIBUTING.md for the full guide, branch rules, and PR workflow.
Branch strategy:main is the only long-lived branch. All work goes through feature branches → PR to main. Branch protection is enforced at repo and org level — no direct pushes, no force pushes, no branch deletion.
# Development mode — load plugin from local directory
claude --plugin-dir ./claude-ops/claude-ops
# Reload after changes
/reload-plugins
See claude-ops/README.md for detailed documentation on each skill, agent, and integration. Full guides, troubleshooting, and the threat model live on the wiki.