No OpenClaw. No cloud agents. No per-token API costs.
Just Claude Code CLI + Claude Max subscription + Docker. That's the whole stack.
I built this over weeks of sleepless nights — a full AI-powered HQ for an MSP company. Lead gen, sales outreach, CRM, contracts, onboarding, finance, government tenders. 12 agents running autonomously on heartbeats.
Then my team never used it. So here it is. Take it.
Screenshots are in Russian — that's how we used it. Fully customizable to any language via the Claude Code prompt below.
| HQ Status (12 agents) | Menu (command center) | Watchdog Alerts |
|---|---|---|
 |
 |
 |
| Live agent status dashboard | 12 buttons, full control | Self-healing Docker restarts |
| Hunter Enrichment | SDR Outreach | Agent Interaction |
|---|---|---|
 |
 |
 |
| ICP scoring, DM verification | SMTP throttle, sequences | Reply to any agent via Telegram |
A complete Paperclip company — real org chart, real agents, real configs that ran a real MSP business in NYC/NJ/PA.
| Doc | What's In It |
|---|---|
COMPANY.md |
Company profile, services, niches, competitive advantages |
DOCUMENTATION.md |
Full architecture: tech stack, workflows, CRM schema, Telegram bot, watchdog |
SHARED-PROTOCOL.md |
The "employee handbook" — company-wide rules all agents follow |
graph TD
CEO["CEO<br/><small>Strategy & Delegation</small>"]
CEO --> SM["Staff Manager<br/><small>HR + Ops</small>"]
CEO --> IC["IT Chef<br/><small>CTO</small>"]
CEO --> FT["Finance Tracker<br/><small>Controller</small>"]
SM --> H["Hunter<br/><small>Lead Gen</small>"]
SM --> SDR["SDR<br/><small>Outbound</small>"]
SM --> CL["Closer<br/><small>Deal Briefs</small>"]
IC --> CM["Contract Manager"]
IC --> PW["Proposal Writer"]
IC --> OA["Onboarding Agent"]
GS["Gov Scout<br/><small>SAM.gov + Tenders</small>"]
LA["Legal Assistant<br/><small>Risk Analysis</small>"]
| Agent | What It Does |
|---|---|
| CEO | Reviews KPIs, delegates tasks, reports to founders. Daily strategist — not an empty coordinator. |
| Hunter | Finds companies with bad IT via job boards, LinkedIn, Apollo.io. ICP scoring 0-100. |
| SDR | Cold emails with branded HTML templates. SMTP throttle: 30s gap, 10 max/heartbeat. |
| Closer | Deal briefs with competitor intel, pricing recs, objection handling. |
| Staff Manager | Monitors all agents, enforces accountability, manages AI org chart. |
| IT Chef | Full CTO replacement when human is offline. Self-heals infrastructure autonomously. |
| Finance Tracker | Invoices, renewals, MRR tracking, payment gap alerts. |
| Contract Manager | Contract expirations, renewal deadlines, compliance. |
| Proposal Writer | SOWs, proposals, NDAs from Word templates. |
| Onboarding Agent | Welcome packages and kickoff scheduling when deals close. |
| Gov Scout | SAM.gov and state portals for IT tenders. |
| Legal Assistant | Contract red flags, liability review. Conservative by design. |
graph TB
subgraph Docker["Docker (9 containers)"]
subgraph Core["Core"]
PS[Paperclip Server]
PU[Paperclip UI]
PG[PostgreSQL]
end
subgraph CRM["CRM Stack"]
TS[Twenty CRM]
TD[CRM Database]
TR[Redis]
end
subgraph Comms["Communications"]
TW[Telegram Webhook]
CS[CRM Sync]
end
subgraph Tunnels["Cloudflare Tunnels"]
T1[CRM]
T2[Telegram]
T3[API]
end
end
subgraph Host["macOS Host"]
WD["Watchdog v2 — 30s polling, auto-restart"]
AG["12 Claude Code Agents — heartbeat-driven"]
BK["Auto-backup — every 6h"]
end
AG -->|claude_local| PS
AG -->|GraphQL| TS
AG -->|MCP| TW
WD -->|monitors| Docker
WD -->|wakes| AG
|
MCP Servers (Tools)
|
Skills (Runtime Knowledge)
|
Telegram Bot — Mobile Command Center (2,200+ lines)
- Single-message UX — entire menu edits ONE message in-place.
editOrSend()cascade: edit caption → edit text → delete + resend. - CRM inside Telegram — 10 views (hot leads, pipeline, nurture, lost, tenders), pagination, lead detail cards with action buttons.
- Pending state machine — tap agent, type message. Persisted to
/tmp/tg-pending-state.jsonwith 10-min auto-expiry. - Voice → Agent — Whisper transcription (base model), transcript becomes task description. Fallback to file attachment.
- Role-based buttons —
ceo_decision(approve/skip lead),account_mgr_call(call tracking),ceo_pricing(pricing approval). - Team-only access — username whitelist. Non-members silently ignored (200 OK).
- Dedup — webhooks deduplicated, but menu/CRM navigation explicitly excluded (fast clicking expected).
Watchdog v2 — Self-Healing Daemon
Bash script via macOS launchd (KeepAlive: true), polls every 30s, 10MB rolling logs.
| Check | Recovery | Time |
|---|---|---|
| Docker Desktop down | open -a Docker |
~120s |
| Paperclip unresponsive | pkill + pnpm dev:once |
~60s |
| Stale DB state | Direct SQL mutations | <10s |
| Twenty CRM down | docker compose up -d |
~15s |
| Telegram webhook down | Restart Node.js | ~2s |
| Cloudflare tunnels dead | Kill orphan + restart | ~4s |
| Hung runs >35min | SQL: mark error, unlock | <5s |
| Agent overdue | POST wakeup | 5min cycle |
caffeinate -si prevents Mac from sleeping. Your AI company works while you sleep.
SHARED-PROTOCOL.md — Corporate Culture for AI
One file defines behavior for ALL agents. Think of it as the employee handbook:
- Task checkout with optimistic concurrency (409 = someone was faster, move on)
- Early exit when idle (save tokens)
- Mandatory Telegram reporting
- Approval gates (first email, proposals, hires, spend >$500)
- Memory protocol (what to remember, what to forget)
- Demand escalation: 0-2h normal → 2h warning → 4-8h alert → 8h+ URGENT
Change one file = change behavior of all 12 agents. No copy-pasting rules.
Hunter — Lead Intelligence Engine
ICP Score = (Fit × 0.40) + (Tech × 0.30) + (Intent × 0.30)
- Fit: employees (20-100 sweet spot), industry (Law/Medical=100, Restaurant=0), geography
- Tech: no MSP + pain signals = 100, strong happy MSP = 0 (auto-skip)
- Intent: "Replace IT provider" job posting = 100, hiring helpdesk = 80, data breach = 75
- Free recon:
recon.sh domain.com→ SSL, DMARC, SPF, WordPress version (zero tokens) - Routing: 80+ HOT (4h SLA) → 60-79 LEAD (24h) → 40-59 Nurture → <40 Skip
SDR — Email Throttling & Sequences
- SMTP throttle: 30s between sends, max 10/heartbeat (IONOS 450 errors taught us this)
- 4-touch sequence: Day 0 (cold), Day 3 (new angle), Day 7 (social proof), Day 14 (breakup)
- Send window: Mon-Thu 8-10 AM ET only
- Reply routing: positive → Closer (1h SLA), pricing → CEO approval, referral → Hunter
- Template enforcement: branded HTML skill must load before EVERY send
IT Chef — Self-Healing CTO
Not an assistant. Full CTO replacement when human is offline.
Auto-fix playbooks (no approval needed): Docker restart, pnpm dev:once, CRM sync restart, pg_resetwal, task unlock, disk cleanup.
Escalation to human only for: SOUL.md changes, new hires, customer data deletion.
Engineering Decisions
| Decision | Why |
|---|---|
dev:once not dev:watch |
Watch mode killed agent processes on file change |
| Hybrid deploy (host + Docker) | Agents need host MCP; CRM needs Docker volumes |
| 409 = no retry | Another agent was faster; move to next task |
| Role-based CRM tools | SDR can't delete leads, Hunter can't touch invoices |
| 3 separate CF tunnels | One down ≠ all down |
| 30s SMTP throttle | IONOS 450 errors killed a whole outreach batch |
| SQL in watchdog | When server is dead, API is dead. Raw PG works. |
| 52% token reduction | Rewrote all instructions. 12 agents × 2-4h = every token counts |
| Mandatory TG reporting | Silent agents get ignored |
| Demand escalation | Polite suggestions don't work. Timed demands do. |
| Cost | One Claude Max subscription ($100/mo) + SMTP/domain |
| Adapter | claude_local — Paperclip talks to Claude Code CLI directly |
| No OpenClaw | No cloud agents, no external runtimes, no per-token API costs |
| Platform | macOS (tested Apple Silicon) + Docker |
- macOS with Docker & Docker Compose
- Paperclip installed
- Claude Code CLI with Max or Pro subscription
- Telegram bot token
- SMTP credentials
- Cloudflare account (optional, for remote access)
# 1. Clone
git clone https://github.com/tr00x/paperclip.git && cd paperclip
# 2. Configure
cp .env.example .env
cp docker/amritech/.env.example docker/amritech/.env
cp twenty-crm/.env.example twenty-crm/.env
# Edit each .env with your credentials
# 3. Start Docker stack
cd docker/amritech && docker-compose up -d
# 4. Start Paperclip
cd ../.. && pnpm install && pnpm dev
# 5. Configure MCP servers
# Edit mcp-servers/mcp-*.json with your API keys- Fork this repo
- Edit
COMPANY.md— goals, team, region - Edit agent
SOUL.mdfiles — personality per agent - Edit
SHARED-PROTOCOL.md— company-wide rules - Adjust
HEARTBEAT.md— wake schedules - Swap MCP servers — your CRM, email, tools
- Update the HTML email template in
skills/amritech-html-email/— agents use this branded template for ALL outbound emails by default. Change the colors, logo, footer, and company name to match your brand.
Open Claude Code in this repo and paste:
I cloned the AmriTech AI HQ template and want to make it mine.
Read amritech-hq/COMPANY.md, amritech-hq/agents/SHARED-PROTOCOL.md, and
amritech-hq/agents/ceo/SOUL.md to understand the current setup.
Then interview me — ask one question at a time:
1. Company name, industry, location?
2. What do you sell and who's your ideal customer?
3. Team members? (name, role, email, Telegram handle each)
4. Big goal? (e.g., "$50k MRR by end of year")
5. Which agents to keep/remove/rename?
6. Any agents to ADD?
7. SMTP provider and credentials?
8. Telegram bot token and chat ID?
9. Need Cloudflare tunnels?
10. Company-wide rules for ALL agents? (goes into SHARED-PROTOCOL.md)
After the interview, update EVERYTHING — COMPANY.md, SHARED-PROTOCOL.md,
every SOUL.md, TOOLS.md, HEARTBEAT.md, .env files, MCP configs, skills.
Delete agents I don't need, create new ones if I added any.
Zero technical knowledge required — just answer the questions.
- Agents need governance, not freedom. Approval gates saved us multiple times.
- Self-healing is non-negotiable. If you don't have auto-restart, you don't have production.
- SOUL + TOOLS + HEARTBEAT = identity + capability + rhythm. Without all three = expensive autocomplete.
- MCP servers are the real power. Agents without tools are useless.
- Heartbeats > always-on. Cheaper, more reliable. Early exit when idle saves tokens.
- Telegram is the control plane. Run the company from your phone.
- Agents that can't demand are useless. Timed escalation actually gets humans to act.
- SHARED-PROTOCOL.md = one file controls all agents. Change once, applies everywhere.
Paperclip is building ClipMart — a marketplace for pre-built AI companies. When it launches, I'll publish full HQ packs there.
| Pack | Agents | Status |
|---|---|---|
| MSP / IT Services (this repo) | 12 agents | Done |
| Marketing Agency | Content, Social, SEO, Campaigns | Planned |
| Dev Shop | PM, QA, Deploy, Sprint | Planned |
| E-commerce | Inventory, Support, Reviews, Pricing | Planned |
| Consulting Firm | Research, Reports, Proposals | Planned |
| Real Estate | Leads, Listings, Follow-ups, Market | Planned |
Published quickly — raw export from production. Works, but not polished.
You might find traces of the original company in some files. I scrubbed what I could, but I'm not spending days sanitizing every corner of a repo nobody uses yet. Nothing sensitive — just internal references.
If there's interest (stars, forks, issues), I'll clean it up properly. Prove me wrong.
| Tool | Purpose |
|---|---|
| Paperclip | AI company orchestration |
| Twenty CRM | Open-source CRM |
| Claude Code | Agent runtime (claude_local) |
| Serena | Semantic code analysis MCP for IT Chef |
| @codefuturist/email-mcp | Gmail SMTP/IMAP for SDR & Onboarding |
| @zhafron/mcp-web-search | Web search for 6 agents |
| Apollo.io | Contact enrichment for Hunter |
| Cloudflare Tunnels | Secure remote access (3 tunnels) |
| Whisper.cpp | Voice-to-task via Telegram |
MIT — do whatever you want with it.
If this saved you time or showed you something new — drop a star.
Stars = motivation. I'll keep building HQ packs if people actually want them.
Built with Paperclip and too many sleepless nights.
By @tr00x