Run Your Own AI Agent Army

Why Self-Host?

Requirements

Quick Start — 5 Minutes

# Recommended: Context 16384, Flash Attention ON, port 1234
curl -s http://localhost:1234/v1/models | head -5

npm install -g openclaw
openclaw init
# Creates: ~/.openclaw/agents/ workspace/ scripts/ logs/

# Get token from @BotFather, then:
openclaw add-provider telegram
openclaw deploy captain --platform telegram
# Done. Message your bot.

openclaw deploy mitch --platform telegram    # Sales
openclaw deploy ghost --platform telegram    # Content
openclaw deploy jobsite --platform telegram  # Invoicing
openclaw fleet status                        # Check fleet

Botwave Desktop — Full Fleet Commander

Manage your entire AI Agent Army from one native window.

Download Botwave Desktop (Beta)

Architecture

The 6-Layer Bash Defense — $0 API Cost

6 layers of defense before a paid token ever gets burned. Bash handles 90% of "AI tasks" free and instant.

# 1. Route by keyword — free, instant, no tokens burned
echo "$msg" | grep -qi "invoice\|quote\|estimate" && ROUTE=jobsite
echo "$msg" | grep -qi "price\|competitor\|market" && ROUTE=scout

# 2. Validate before calling — don't waste a call on empty input
[ -z "$msg" ] && echo "Empty input" && exit 0

# 3. Cache with TTL — same question within 24h = $0
HASH=$(echo "$prompt" | md5sum | cut -d' ' -f1)
if [ -f "cache/$HASH.json" ] && [ ! $(find "cache/$HASH.json" -mmin +1440) ]; then
  cat "cache/$HASH.json" && exit 0
fi

# 4. Compress context before inference — 10K tokens → 500
echo "$convo" | jq '.messages[-5:]'

# 5. Only extract what you need from the response
curl -s localhost:1234/v1/chat/completions -d "$payload" | jq '.choices[0].message.content'

# 6. Store result in cache for next time
echo "$response" | tee "cache/$HASH.json"
# GPT-4 cost: $0.15/interaction → Botwave: $0.000

Multi-Model Routing

# FREE ONLY (high-volume)
Captain  → LM Studio      $0   Ghost → Groq (Llama 3.3) $0
Scribe   → LM Studio      $0   Scout → Groq             $0
Content  → LM Studio      $0   Oracle → OpenRouter/free  $0

# FREE + PAID FALLBACK (revenue-generating)
Mitch    → Groq → xAI Grok-2         (paid only for deal closing)
JobSite  → LM Studio → Mistral Large  (paid only for invoices)
Closer   → Groq → OpenRouter/paid     (revenue = worth it)

# LOCAL ONLY (security/privacy — NEVER cloud)
DickTracy → LM Studio (mandatory)  $0
HOMIIEE   → LM Studio (mandatory)  $0

Agent Specializations

Cron Automation

# 18+ jobs running right now:
# Every 6h  → Ghost posts to Twitter
# Daily     → Mitch checks leads, sends follow-ups
# Nightly   → Scout scrapes competitor pricing
# Weekly    → Oracle market analysis report

openclaw cron add ghost "Post content" --every 6h
openclaw cron add mitch "Check leads" --every 1d
openclaw cron list

Full Audit Trail

~/.openclaw/logs/
  commands.log    # Every bot command
  fleet.log       # Agent orchestration
  revenue.log     # Payment events
  engage.log      # User interactions

~/.openclaw/agents/*/sessions/*.jsonl
  # Full conversation history — every msg, tool call, response

Script Factory

Every automation saved as reusable bash. Build massive codebases with near-zero paid API cost.

~/.openclaw/scripts/
  phase7-dicktracy-overhaul.sh   # Agent fixes
  phase8-api-strategy.sh         # Model routing
  phase10-business-overhaul.sh   # THIS script
  ai-router-config.json          # Router config
  marketing/                     # Launch content
  pentest/                       # 80+ security scripts

Security (Boss + Operator)

24/7 Headless

tmux new-session -d -s botwave
tmux send-keys "lms server start --port 1234 --ctx 16384" Enter
tmux send-keys "openclaw start --all-agents" Enter
# Detach. Your agents run. Check from Telegram anytime.

Roadmap

• Now: 15 agents, Telegram-first, 6-layer bash defense, local routing, Desktop app (Beta)
• Q2 2026: ClawHub skill marketplace
• Q3 2026: Raspberry Pi images (BotWave Node)
• Q4 2026: Docker containerized agents + seccomp
• 2027: Mobile companion, voice control

FAQ

Do I need a GPU?

No. CPU is fine for 7B on 8GB+.

Can I use my own models?

Yes. LM Studio, Ollama, any OpenAI-compatible endpoint.

Is data really private?

Self-hosted = your machine. No telemetry, no phone-home.

How is this different from CrewAI/n8n?

Those are frameworks. Botwave is a complete agent OS: phone control, zero code, zero API, full audit trail.