A guy in one of my Facebook groups got banned from Claude last week.
Two months of heavy use. Over $10,000 in credits. Nine software products built on top of it. Gone at midnight on a Saturday. Mid-session.
"Your account has been disabled after an automatic review of your recent activities."
No warning. No email. Just locked out.
That post hit different for me — because I had been building the same way. My entire operation for Load Bearing Empire (6 AI-powered businesses I run out of Houston) was sitting on managed platforms. Twilio for calls. Railway for hosting. VAPI for voice agents. Vercel for everything else. One automated policy check away from losing it all.
So I built CLAWKILLER.
What CLAWKILLER Is
CLAWKILLER is a production-grade, self-hosted multi-agent autonomous operations platform. It takes everything OpenClaw offers and extends it with enterprise-grade infrastructure, multi-agent orchestration, persistent memory, job scheduling, and business-specific tooling — running entirely on infrastructure I own.
Here is what it does in plain terms: I send a Telegram message that says "create an Instagram page for Quiet Hours Valet." CLAWKILLER routes that through a supervisor agent, spins up a Playwright browser on my VPS, fills out the signup form, handles phone verification automatically using my Asterisk PBX to capture the SMS code, and sends me a screenshot when the page is live.
No VAPI. No Twilio. No OpenClaw. Just Python, Node.js, and a $6/month Vultr VPS I already owned.
The Architecture
Telegram / Web App (LB Agent OS)
|
Node Gateway (port 3000)
|
FastAPI Gateway (port 8000)
|
LangGraph Supervisor (Claude Sonnet)
|
BROWSER COMMS CODE BUSINESS RESEARCH FILE SCHEDULER
AGENT AGENT AGENT AGENT AGENT AGENT AGENT
(GPT-4o)(Gemini)(Sonnet)(Sonnet) (Gemini) (GPT) (Gemini)
|
Playwright + Claude Vision Recovery
|
Asterisk PBX (SMS capture on VoIP.ms DIDs)
|
Supabase (sessions, memory, CRM, pgvector)
|
BullMQ + Redis (job queues)
The supervisor is the brain. It reads every incoming message and decides which specialist agent handles it. Each agent has a defined tool set — the browser agent cannot send SMS, the comms agent cannot execute shell commands. Containment by design.
The Seven Specialist Agents
Supervisor (Claude Sonnet) routes every command. Loops back after each agent completes to decide if more work is needed. Hard-capped at 10 iterations to prevent infinite loops.
Browser Agent (GPT-4o + Playwright) navigates websites, fills forms, creates social media pages. When it hits an unexpected UI state, it takes a screenshot and asks Claude vision what to do next. This recovery loop is what makes it actually work instead of crashing.
Comms Agent (Gemini Flash) routes all outbound communication through my Asterisk PBX. SMS goes out over VoIP.ms SIP. When a social platform sends an SMS verification code to one of my VoIP.ms DIDs, Asterisk captures it via AGI script, POSTs it to the FastAPI gateway, which injects it directly into the waiting browser task. The whole thing resumes automatically.
Code Agent (Claude Sonnet) executes Python and shell in a sandboxed directory with a blocked command list.
Research Agent (Gemini Flash) handles web search, property lookups via existing CrawlOS Edge Functions, company intelligence.
Business Agent (Claude Sonnet) runs full CRM operations against Supabase — lead creation, scoring, pipeline updates, job orders, invoice queuing.
Scheduler Agent (Gemini Flash) manages BullMQ job queues across five queues: leads, comms, billing, scrape, dispatch.
The SMS Interception Flow
Each of my six businesses has its own VoIP.ms DID on my Asterisk server. When the browser agent hits a phone verification screen, it marks the task awaiting_sms and records which DID it used. The platform texts that DID. Asterisk captures the SIP MESSAGE, an AGI Python script extracts the code and POSTs it to the FastAPI gateway. The gateway injects it into the waiting task. Browser agent enters the code and continues.
Zero human involvement past the initial command.
The Memory System
Three tiers:
Working Memory (Redis + LangGraph checkpointer) — current session state. If a task fails midway, LangGraph resumes from the last checkpoint.
Episodic Memory (Supabase Postgres) — every conversation turn, tool execution, and routing decision logged. Query exactly what any agent did for any task.
Semantic Memory (Supabase pgvector) — content embedded via OpenAI text-embedding-3-small with cosine similarity search. Any agent can pull relevant context from past sessions.
Cost Breakdown
| Service | Monthly |
|---|---|
| Vultr VPS | $6 |
| Supabase | $0 |
| Redis (self-hosted) | $0 |
| VoIP.ms 6 DIDs | ~$12 |
| Anthropic API (hard-capped) | $30-50 |
| OpenAI API | $10-20 |
| Google Gemini Flash | $2-5 |
| Total | ~$60-90/mo |
The spec I built from estimated OpenClaw + managed infra at $350-$1,000/month. The other number that matters: zero subscriptions cancellable by someone else's automated review system.
What I Skipped On Purpose
Railway — already had a Vultr VPS with Asterisk. Three systemd services replace Railway.
Twilio — Asterisk + VoIP.ms handles everything. The comms agent writes Asterisk call files directly.
OpenClaw's plugin marketplace — Cisco found data exfiltration in a third-party OpenClaw skill. No marketplace, no problem.
The Real Point
Most people reading that Facebook ban post suggested having backup accounts across multiple platforms.
That is the wrong frame. The right answer is not better redundancy across platforms you do not control. The right answer is to control the platform.
Every piece of infrastructure in this stack — the VPS, the PBX, the database, the agents, the memory system, the job queues — is something I own and operate. No one can ban me from my own server. No automated review system can disable my Asterisk instance. No terms of service violation can kill my Redis queue.
That is operational sovereignty. That is what CLAWKILLER is built on.
Building Load Bearing Empire in public. Follow: @DomoniqueLuchin
Top comments (0)