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.

10.0 KB
Raw Blame

Architecture

claude-ops is a Claude Code plugin (v1.7.0) that installs as a business operating system on top of the Claude Code CLI. It exposes 30 slash-command skills that orchestrate 14 autonomous agents, a persistent background daemon, and a bundled Telegram MCP server. The primary inputs are tool calls from Claude Code sessions; the primary outputs are structured reports, drafted communications, and autonomous code/infrastructure actions executed via shell and API calls.


Component Diagram

graph TD
    User["User (Claude Code session)"]
    Skills["Skills Layer\n30 slash commands\nskills/<name>/SKILL.md"]
    Agents["Agents Layer\n14 autonomous agents\nagents/<name>.md"]
    Daemon["ops-daemon\nlaunchd-managed\nscripts/ops-daemon.sh"]
    Hooks["Hooks\nhooks/hooks.json"]
    Bin["Bin Scripts\nbin/ops-*"]
    TelegramMCP["Telegram MCP Server\ntelegram-server/index.js"]
    Registry["Project Registry\nscripts/registry.json"]
    Cache["Daemon Cache\n~/.claude/plugins/data/.../daemon-cache.json"]
    Prefs["User Preferences\nplugin.json userConfig"]
    Secrets["Secrets (Doppler / 1Password / env)"]

    User -->|slash command| Skills
    Skills -->|spawn| Agents
    Skills -->|read| Registry
    Skills -->|read| Cache
    Skills -->|call| Bin
    Agents -->|Bash + tool calls| Bin
    Daemon -->|writes| Cache
    Daemon -->|runs| Bin
    Hooks -->|PreToolUse / SessionStart / Stop| Skills
    TelegramMCP -->|MCP protocol| Skills
    Prefs -->|userConfig injection| Skills
    Secrets -->|env vars| Bin
    Secrets -->|env vars| TelegramMCP

Data Flow

A typical /ops:go morning briefing flows as follows:

  1. SessionStart hook — hooks/hooks.json fires bin/ops-welcome (displays setup status) and scripts/setup.sh (surfaces any unresolved configuration issues).
  2. Skill invocation — the user runs /ops:go; the skills/ops-go/SKILL.md skill is loaded by Claude Code.
  3. Cache read — the skill calls bin/ops-gather (or reads the pre-warmed daemon-cache.json written by the daemon's briefing-pre-warm service every 2 minutes).
  4. Agent fan-out (optional) — for deeper analysis (/ops:yolo), the skill spawns 4 parallel agents (CEO, CTO, CFO, COO) via the Claude Agent SDK. Each agent runs independently using the session-default model and writes its analysis to a temp file.
  5. Merge + present — the main skill orchestrator reads all agent output files, merges them into a unified report, and returns it to the user in the Claude Code session.
  6. Stop hook — bin/ops-post-session-cleanup runs on session end to clean up temp files and stale PIDs.

For communication skills (/ops:inbox, /ops:comms), a PreToolUse hook (bin/ops-pretool-wacli-health) checks the WhatsApp daemon health file before any wacli Bash call and surfaces a warning if the sync is stale (>10 min).


Key Abstractions

Abstraction Location Description
plugin.json claude-ops/.claude-plugin/plugin.json Plugin manifest: name, version, userConfig schema (20 configurable keys), MCP server declarations
SKILL.md claude-ops/skills/<name>/SKILL.md Each skill is a markdown prompt file loaded by Claude Code as a slash command
Agent .md files claude-ops/agents/<name>.md Each agent is a markdown prompt file with frontmatter declaring resource settings such as effort, maxTurns, and memory scope; model is optional
hooks.json claude-ops/hooks/hooks.json Declares SessionStart, PreToolUse (Bash matcher), and Stop lifecycle hooks
registry.json claude-ops/scripts/registry.json User-maintained project index with alias, repo, infra platform, revenue model, and GSD flag per project
ops-daemon.sh claude-ops/scripts/ops-daemon.sh Shell supervisor for 7 persistent background services; managed via launchd
daemon-cache.json ~/.claude/plugins/data/.../daemon-cache.json Pre-warmed briefing data written by briefing-pre-warm every 2 min; read by /ops:go for <3s load
CLAUDE.md claude-ops/CLAUDE.md Plugin-root rules (Rules 0–5) that override individual skill instructions for all agents
Telegram MCP server claude-ops/telegram-server/index.js Node.js MCP server providing MTProto user-auth Telegram access (personal account, not bot API)

Directory Structure Rationale

claude-ops/
├── .claude-plugin/
│   └── plugin.json        # Plugin manifest and userConfig schema
├── agents/                # 12 agent prompt files (.md) — spawned by skills
├── bin/                   # 22 shell/node scripts (ops-*) — the executable layer
├── docs/                  # Reference documentation for plugin internals
├── hooks/
│   └── hooks.json         # Lifecycle hook declarations (SessionStart, PreToolUse, Stop)
├── output-styles/         # Output formatting guides used by skills
├── scripts/               # Daemon, cron jobs, setup wizard, project registry
├── skills/                # 22 skill directories, each containing a SKILL.md prompt
├── telegram-server/       # Bundled MCP server for Telegram MTProto access
├── tests/                 # Bash test suite (no-secrets, skills-lint, hooks, bin-scripts)
├── CHANGELOG.md           # Version history
├── CLAUDE.md              # Plugin-root rules applied to all skills and agents
├── LICENSE                # MIT
└── package.json           # Runtime deps for bin .mjs scripts (telegram, playwright)
  • skills/ — User-facing layer. Each subdirectory is one slash command. Skills are markdown prompt files; they contain no executable code, only instructions for Claude.
  • agents/ — Background worker layer. Agent .md files are spawned by skills using the Claude Agent SDK. They can declare resource budgets (maxTurns, effort) and memory scopes; when model is omitted they inherit the session default.
  • bin/ — Executable layer. Shell and Node.js scripts that interact with external CLIs (wacli, aws, gh, doppler, gog) and APIs. Skills and agents call these via Bash tool calls.
  • scripts/ — Infrastructure layer. The daemon supervisor, launchd plists, cron scripts, and the project registry live here.
  • telegram-server/ — Bundled Telegram user-auth server (MTProto, not a bot). Not auto-started from .mcp.json (that file is empty so sessions do not spawn a stdio MCP each). Started on demand via /ops:setup / host config.
  • hooks/ — Automation layer. Three lifecycle hooks: SessionStart (welcome + setup check), PreToolUse/Bash (WhatsApp health check), and Stop (cleanup).
  • tests/ — Validation layer. Bash scripts that check for committed secrets, skill markdown lint, hook schema validity, and bin script executability. Run via tests/run-all.sh.

Daemon Architecture

The ops-daemon is a launchd-managed supervisor that runs 7 background services continuously:

Service Cadence Purpose
briefing-pre-warm every 2 min Runs bin/ops-gather to keep /ops:go cache hot
wacli-sync persistent Keeps WhatsApp connected via wacli, auto-reconnects
memory-extractor every 30 min Spawns memory-extractor agent to refresh memories/
inbox-digest every 15 min Pre-classifies communications across all channels
store-health every 10 min Shopify orders and inventory polling (if configured)
competitor-intel hourly Background market and competitor monitoring
message-listener persistent Surfaces urgent patterns ("urgent", "ASAP", "fire", "down")

The daemon installs at setup Step 2c — deliberately early — so the briefing cache is pre-warmed while the user completes the remaining integration configuration steps.


Agent Model Assignments

Agent class Model Use case
C-suite analysts (yolo-ceo, yolo-cto, yolo-cfo, yolo-coo) session default (no pin) High-depth strategic, technical, financial, operational analysis
Scanner, fix, and daemon agents session default (no pin) Efficient read-scan-report loops and targeted code fixes
memory-extractor session default (no pin) High-frequency (every 30 min) low-cost contact extraction

Plugin Rule Enforcement

claude-ops/CLAUDE.md defines 6 rules (Rule 0 through Rule 5) that apply to every skill and agent in the plugin, overriding any conflicting instruction in individual files:

  • Rule 0 — No personal data ever committed (public open-source repo)
  • Rule 1 — Maximum 4 options per AskUserQuestion call
  • Rule 2 — Never delegate commands to the user; use Bash tool instead
  • Rule 3 — Never auto-skip channels or integrations during setup
  • Rule 4 — Background by default during setup and configuration flows
  • Rule 5 — Destructive actions require explicit per-action confirmation