The LocalClaw
Playbook
A step-by-step guide to deploying autonomous AI agents for your local business using OpenClaw and NVIDIA NemoClaw. No technical degree required.
What's Inside
Everything you need — from understanding what AI agents are, to deploying a fully secured, multi-channel agent that works for your business 24 hours a day.
Why Local Businesses
Need AI Agents Now
Your competitor three blocks away is about to deploy an AI agent that answers every lead in under 60 seconds, books appointments at 2am, and never takes a day off. Here's what that means for you.
48% of small business inquiries go unanswered within the first hour. That's the hour where the customer decides who to call next.
The local business landscape has shifted permanently. The question is no longer “should I use AI?” — it's “how fast can I deploy before my competitor does?”
24/7 Availability
Customers contact businesses at all hours. An AI agent never sleeps, never misses a message, and responds in seconds — not hours.
Instant Lead Response
Speed to lead is everything. Responding within 5 minutes increases conversion by 900% vs responding in 30 minutes.
Automated Booking
No more back-and-forth. The agent checks your calendar, proposes slots, confirms the booking — fully automatically.
Lower Operating Cost
A single AI agent handles what would normally require a part-time hire — at a fraction of the monthly cost.
Multi-Channel Presence
One agent across WhatsApp, Gmail, Telegram, and Instagram DMs simultaneously. Stay present everywhere your customers are.
Enterprise Security
With NemoClaw, your customer data stays protected. Policy-based controls ensure your agent does only what you allow.
Local business owners, operations managers, and agency owners who want to deploy AI agents. You do not need to know how to code. Every step in this guide is designed for non-technical people.
What Is OpenClaw?
The Agent Engine
OpenClaw is an open-source, self-hosted gateway that connects your messaging apps — WhatsApp, Telegram, Gmail, Discord — to an always-available AI agent. Think of it as the engine under the hood.
Multi-Channel Gateway
A single Gateway process serves WhatsApp, Telegram, Discord, and iMessage simultaneously from one install.
Persistent Memory
Your agent remembers past conversations, customer names, preferences, and history — per session and across sessions.
Tool Use
Agents can browse the web, read files, call APIs, book calendar slots, and execute code — not just chat.
Multi-Agent Routing
Route different requests to specialized agents. One for bookings, one for support, one for sales — all coordinated.
OpenClaw is free and open source. You own your data, your agent, and your infrastructure. No vendor lock-in. 329,000+ GitHub stars.
What Is NemoClaw?
The Security Shield
NemoClaw is NVIDIA's open-source security wrapper for OpenClaw. It adds a policy enforcement layer — called OpenShell — that controls exactly what your agent can see, do, and send.
When you give an AI agent access to your inbox, calendar, and files — you're giving it the same permissions as a full-time employee. Without NemoClaw, there are zero guardrails on what it accesses or where your data goes.
| FEATURE | RAW OPENCLAW | OPENCLAW + NEMOCLAW |
|---|---|---|
| Policy-based access control | ✗ None | ✓ OpenShell enforced |
| File system scope limits | ✗ Full system access | ✓ Directory-scoped per policy |
| Credential protection (.env) | ✗ High risk | |
| Data routing (local vs cloud) | ✗ Cloud by default | ✓ Local Nemotron first |
| Prompt injection protection | ✗ None | ✓ Agent scope limited |
| Audit logging | ✗ None by default | ✓ Full agent action log |
| External API call controls | ✗ Unrestricted | ✓ Allowlist via OpenShell |
| License | MIT | Apache 2.0 |
The LocalClaw Stack
How It All Fits Together
LocalClaw combines OpenClaw (the engine) and NemoClaw (the shield) into a fully managed, secure agent deployment for your business. Here's the complete picture.
You don't have to manage any of this yourself. LocalClaw deploys, configures, and maintains the entire stack for you — same day. You get the dashboard, the agent, and the security without touching a single line of code.
Pre-Installation
Checklist
Before you install anything, run through this checklist. Most failures happen because of one skipped prerequisite.
System Requirements
- Node.js 22 LTS or higher — Check with
node -vin your terminal. Download at nodejs.org if needed. - npm 9 or higher — Comes bundled with Node. Check with
npm -v. - 4GB RAM minimum — 8GB recommended if you plan to run local Nemotron routing.
- Stable internet connection — Required for initial install and cloud AI model calls.
- Administrator / sudo access — Required to install the OpenClaw background daemon.
Accounts & API Keys
- AI Provider API Key — Get one from OpenAI (platform.openai.com) or Anthropic (console.anthropic.com). GPT-4o or Claude 3.5 Sonnet recommended.
- Google Account (recommended) — For Gmail and Google Calendar channel integration.
- WhatsApp Business Number (optional) — Requires a Meta Business account and a dedicated phone number.
- Telegram Account (optional) — Fastest channel to connect. 5 minutes via BotFather.
For 24/7 Deployment (Recommended)
- A VPS or Cloud Server — DigitalOcean Droplet, Hetzner Cloud, or Vultr. $6–$12/month keeps your agent running even when your laptop is off.
- SSH Access to your server — You'll need the IP address and login credentials.
- Domain name (optional) — For a clean webhook URL when connecting WhatsApp and Instagram channels.
If this checklist feels like a lot — that's exactly why LocalClaw exists. We handle every single item on this list as part of our deployment service. Book a free 15-minute call and your agent will be live today.
Installing OpenClaw
Mac · Windows · Linux
OpenClaw installs as a global npm package. One command, any operating system. The whole process takes under 2 minutes.
Step 1 — Install OpenClaw globally
npm install -g openclaw@latestStep 2 — Verify the install
openclaw --version # Expected: openclaw/x.x.x node/v22.x.x linux/x64
Step 3 — Preview the Dashboard
openclaw dashboard # Opens http://127.0.0.1:18789/ in your browser
Run your terminal as Administrator on Windows. Right-click your terminal app and choose “Run as administrator” before running the install command. This is required for the global install and daemon setup.
If you're SSH'd into a remote server, run all commands as a non-root user with sudo privileges. Do not run as root directly — the daemon install needs a proper user context.
Running Your First
Onboarding
The openclaw onboard command walks you through the entire setup with a guided terminal wizard — AI provider, model selection, channels, and daemon installation. Takes about 5 minutes.
openclaw onboard --install-daemonWhat the wizard walks you through
- 1
Choose your AI provider
Select OpenAI, Anthropic, Mistral, or Groq. Paste your API key when prompted. Stored locally in
~/.openclaw/openclaw.json— never sent anywhere else. - 2
Select your model
Recommended:
gpt-4oorclaude-3-5-sonnet. For cost-sensitive deployments,gpt-4o-minihandles routine tasks well at a much lower cost per token. - 3
Choose channels to connect
Select the channels you want. You can always add or remove channels later from the dashboard. Telegram is the fastest to connect — 5 minutes via BotFather.
- 4
Install the background daemon
The
--install-daemonflag installs OpenClaw as a system service. It starts automatically when your server boots — essential for 24/7 operation without manual restarts. - 5
Open the Control Dashboard
OpenClaw opens
http://127.0.0.1:18789/in your browser. You'll see the Gateway status, connected channels, and live session feed. Your engine is running.
When the dashboard shows a green “Gateway: Running” indicator — your agent engine is live. Do not connect live channels yet. Install NemoClaw security first (next chapter).
Installing NemoClaw
The Security Layer
NemoClaw wraps your OpenClaw install with NVIDIA's OpenShell policy engine. Install it before connecting any live channels. This is your agent's firewall.
Install NemoClaw before you connect WhatsApp, Gmail, or any live channel. Once real customer data flows through your agent, you want OpenShell policies active from day one — not as an afterthought.
Fresh Install (new OpenClaw setup)
# Step 1: Download and install NemoClaw curl -fsSL https://nvidia.com/nemoclaw.sh | bash # Step 2: Run interactive onboarding nemoclaw onboard # Step 3: Confirm everything is active nemoclaw status
Migrating an Existing OpenClaw Setup
# Import your existing OpenClaw config automatically nemoclaw onboard --import-from=openclaw
Enable Local AI Routing (NVIDIA GPU only)
# Route sensitive workloads to on-device Nemotron model nemoclaw config set router.mode=local-first # Verify routing mode nemoclaw config get router.mode
nemoclaw status should show: OpenShell: active · Policy: loaded · Audit log: enabled. If all three are confirmed — your security layer is live and protecting your stack.
OpenShell Policy Setup
Protecting Your Business
OpenShell policies are YAML files that define the exact boundaries of what your agent is allowed to do. Every action your agent attempts is checked against this file before it executes. If it's not in the policy — it doesn't happen.
The Starter Business Policy
Save this as ~/.openclaw/openshell.yaml. Customize the file paths and approved domains to match your business setup. This is a safe, restrictive baseline for any local business.
version: "2026-03"
agent_id: "my-business-agent"
capabilities:
file_system:
allow:
- "/home/user/business/contacts/**" # CRM / customer data
- "/home/user/business/bookings/**" # Appointment files
- "/tmp/agent-workspace" # Temporary scratch space
deny:
- "**/.*" # ALL hidden files (.env, .ssh, etc.)
- "/home/user/.ssh/**" # SSH keys — explicit block
- "/etc/passwd" # System credential files
network:
mode: "allowlist"
allow_domains:
- "api.openai.com"
- "api.anthropic.com"
- "calendar.googleapis.com"
- "gmail.googleapis.com"
- "api.telegram.org"
deny_all_else: true
local_routing:
force_local_on_sensitive: true
logging:
level: verbose
destination: "/var/log/openshell/agent.log"The "**/.*" deny pattern is your most powerful security rule. One wildcard blocks every hidden file on the system — your .env, .ssh, browser session files, cached credentials. Never remove it.
Reload Policy Without Restarting
# Validate your policy file for YAML syntax errors nemoclaw policy validate # Apply changes live — no daemon restart needed nemoclaw policy reload
Connecting Your
Channels
OpenClaw supports every channel your customers use. Connect one or all of them — your agent handles them all through the same Gateway with a shared memory and context.
Telegram — Fastest Setup (5 Minutes)
- 1
Open Telegram and search for @BotFather
BotFather is Telegram's official bot creation service. Start a chat with it — it's a verified blue-check account.
- 2
Send /newbot
BotFather asks for a display name (e.g. “Maya from Mike's Barbershop”) and a username ending in
bot(e.g.@mikesbarbershop_bot). - 3
Copy the bot token
BotFather gives you a token like
7845123456:AAFxxx.... This is your channel key — treat it like a password. - 4
Add token to OpenClaw
Run
openclaw channels add telegramand paste the token when prompted. Or add it directly in the dashboard under Channels → Telegram. - 5
Send a test message
Open your bot in Telegram and say “Hello”. You should get an agent response within 2–3 seconds.
WhatsApp — Meta Business API
- 1
Create a Meta Business account
Go to business.facebook.com and set up a free Business Account if you don't have one. Takes 3 minutes.
- 2
Set up WhatsApp Business API
In Meta Business Suite → WhatsApp → Get Started. Register a dedicated phone number for your business agent.
- 3
Copy your credentials
From the Meta Developer portal, copy your Phone Number ID, Business Account ID, and System User Access Token.
- 4
Connect via OpenClaw
Run
openclaw channels add whatsappand paste your credentials. OpenClaw registers the webhook with Meta automatically.
Run openclaw channels add gmail. You'll be redirected to Google OAuth — authorize access, and your agent can read inbound emails, draft replies, and manage calendar bookings immediately. Add calendar.googleapis.com and gmail.googleapis.com to your OpenShell allowlist first.
Building Your First
Business Agent
Every OpenClaw agent is defined by a SOUL.md file — a plain text document describing who your agent is, what it knows, and exactly how it should behave. No coding required. The more detailed, the better.
Think of it as the job description, personality profile, and training manual for your AI employee — all in one document. You write it in plain English. OpenClaw reads it and your agent becomes that person.
The Complete SOUL.md Template
# IDENTITY Name: Maya Business: Mike's Barbershop — Atlanta, GA Role: Customer Service Agent & Booking Assistant Tone: Friendly, warm, professional. Never robotic. Like a real person. Language: English. Simple, clear sentences. No jargon. # WHAT YOU KNOW Services: - Haircut: $35 · 45 minutes - Beard Trim: $20 · 20 minutes - Haircut + Beard: $50 · 60 minutes - Kids Cut (under 12): $25 · 30 minutes Hours: Monday–Saturday 9am–7pm · Sunday CLOSED Address: 123 Main Street, Atlanta, GA 30301 Phone: (404) 555-0192 Parking: Free street parking on Main Street # HOW TO HANDLE REQUESTS Booking request: Ask: which service? Preferred date and time? Check Google Calendar for open slots. Propose 2-3 options if requested slot is taken. Confirm: date, time, service, address, barber (if specified). Send reminder 24h before via same channel. Pricing questions: Answer directly from the services list above. Never say "I'm not sure" — always give a clear answer. Complaints: Acknowledge. Apologize genuinely. Offer a solution (discount or redo). Escalate to Mike if the customer is very upset or requests a refund. New leads (first contact): Greet warmly. Ask what brings them in. Offer to book right away. Do not make them ask twice. # CONVERSATION STYLE Opening: "Hey! Welcome to Mike's Barbershop. How can I help you today?" Closing: "See you soon at Mike's! 💈" Keep responses: Short and conversational. 1-3 sentences per message. # WHAT YOU NEVER DO - Never promise a specific barber unless Mike confirms availability. - Never discuss competitor pricing or badmouth other shops. - Never share customer data, booking details, or contact info. - Never make promises about wait times you can't verify.
Register and Deploy Your Agent
# Register the agent with OpenClaw openclaw agents add --soul ./SOUL.md --name "maya" # Assign it to your connected channels openclaw agents assign maya --channels telegram,gmail # Test it interactively in the terminal before going live openclaw chat --agent maya # Check agent status openclaw agents list
Real-World Agent
Recipes
Five ready-to-deploy agent templates for the most common local business use cases. Copy the SOUL.md core, fill in your business details, and go live in minutes.
Triages every email, drafts replies, flags urgent messages for you
Monitors your Gmail inbox and categorizes every incoming email — new lead, support request, complaint, or junk. Auto-replies to routine inquiries and queues complex ones for your review with a draft already written.
Role: Email Triage & Response Agent Priority 1: New leads — reply within 2 min with intro + booking link Priority 2: Support requests — acknowledge + estimated resolution time Priority 3: Routine FAQ — auto-reply with direct answer Flag for human: Complaints, refunds, legal language, angry tone Never: Send invoices, make refund promises, share customer data
Books, reschedules, and confirms appointments via any channel
Reads your Google Calendar in real time, proposes available slots, collects customer details, creates the calendar event, and sends a confirmation message — all without any manual input from you.
Role: Appointment Scheduling Agent Tools: Google Calendar (read + write enabled) Booking flow: Ask service type → check availability → propose 3 slots → confirm Confirmation: Date, time, service, address, cancellation policy Reminders: Send 24h before via same channel customer used to book Reschedule: Accept if requested 4+ hours before the appointment No-shows: Follow up 15 minutes after missed appointment time
Scores inbound leads and routes hot ones to you instantly
When a new inquiry arrives on any channel, this agent asks 3–5 qualifying questions, scores the lead (hot/warm/cold), and either books them directly or sends you an instant Telegram alert with a full summary.
Role: Lead Qualification & Routing Agent Questions: Budget range? Timeline? Specific service? Location? Heard of us how? Hot lead: Ready to buy + budget fits → book immediately, alert owner Warm lead: Interested, not urgent → add to follow-up, send useful content Cold lead: Browsing → send FAQ + “book when you're ready” message Owner alert:Telegram notification for every hot lead with full context
Writes and schedules on-brand content for Instagram, Facebook & X
Generates on-brand posts, captions, and hashtag sets based on your business niche. Schedules them at optimal posting times and maintains a consistent content calendar with zero manual effort.
Role: Social Media Content & Scheduling Agent Brand voice: Professional, local, community-first. No slang. No politics. Post types: Before/after, tips, promotions, testimonials, behind-the-scenes Frequency: 1 post/day Instagram · 3 tweets/day · 3x/week Facebook CTA: Always include booking link or phone number at end of post Hashtags: 5–10 per post, mix of niche + local + broad Never post: Politics, religion, competitor comparisons, unverified claims
Handles FAQs, complaints, and follow-ups 24 hours a day
Answers every customer question using your business knowledge base. Handles complaints with empathy, resolves simple issues automatically, and escalates complex cases to you with full conversation context already attached.
Role: Customer Support Agent Tone: Empathetic, patient, solution-focused. Never defensive. FAQs: Answer directly and confidently from your knowledge base Complaints: Acknowledge → apologize → offer solution → follow up in 24h Escalate: Legal language, refunds over $100, third repeat complaint CSAT: Ask for 1–5 star rating after every resolved issue Log: Tag every conversation with type: FAQ / complaint / escalated
The Privacy Router
Keeping Sensitive Data Local
NemoClaw's Privacy Router evaluates every agent request and decides in real time whether to process it on a local NVIDIA Nemotron model or route it to a cloud AI. Your data never goes where it shouldn't.
What Stays Local
Customer names, phone numbers, emails, payment references, booking history, and any message containing personal details — all processed on-device.
What Goes to Cloud
General content writing, knowledge lookups, planning tasks, and anything not containing sensitive data — handled by cloud models for best quality.
No GPU? No Problem
Without an NVIDIA GPU, NemoClaw sanitizes prompts before cloud routing — stripping PII before the request leaves your machine.
Full Audit Trail
Every routing decision is logged — what was sent where, what was blocked, and which policy rule triggered. Full compliance visibility.
With local routing enabled, customer PII never leaves your server. This is a significant advantage for businesses in regulated environments or those serving customers in the EU.
Monitoring
& Audit Logs
Every action your agent takes is logged. This is how you see what it's doing, catch problems early, and stay in complete control of an autonomous system running on your behalf.
Viewing Live Logs
# Watch the OpenShell audit log in real time tail -f /var/log/openshell/agent.log # View OpenClaw session logs (last 100 entries) openclaw logs --tail 100 # Filter by specific channel openclaw logs --channel telegram --tail 50 openclaw logs --channel gmail --tail 50 # Check overall health nemoclaw status openclaw agents list
Timestamp
Exact time of every agent action. Useful for debugging response delays and compliance reporting.
Action Type
What the agent did — file read, API call, email sent, calendar checked, message replied.
Policy Result
Whether OpenShell allowed or blocked the action, and which specific rule applied.
Data Routing
Whether the request processed locally on Nemotron or went to a cloud model.
Pipe your OpenShell log to a simple script that Telegrams you when it sees DENY events. If your agent is getting blocked repeatedly, it means your policy needs updating — or someone is probing it.
#!/bin/bash
# Watch for DENY events and send a Telegram alert
tail -f /var/log/openshell/agent.log | grep --line-buffered "DENY" | while read line; do
curl -s -X POST "https://api.telegram.org/bot${BOT_TOKEN}/sendMessage" \
-d chat_id="${YOUR_CHAT_ID}" \
-d text="OpenShell DENY: ${line}"
doneTroubleshooting
Common Issues
The most common problems and their exact fixes. Most issues resolve in under 5 minutes. Run through these in order before reaching out for support.
Run openclaw status. If it shows “stopped”, run openclaw start. If the daemon is running but messages aren't coming through, check your channel token — WhatsApp tokens expire if not refreshed. Run openclaw channels refresh whatsapp.
Your openshell.yaml file has a syntax error. Run nemoclaw policy validate to find it. The most common cause is inconsistent indentation — YAML requires exact 2-space indentation throughout.
Add the path to your openshell.yaml allow list under capabilities.file_system.allow. Save the file, then run nemoclaw policy reload — no restart needed. The change takes effect immediately.
OAuth tokens expire every 7 days without refresh. Run openclaw channels refresh gmail to force a fresh token. Also confirm calendar.googleapis.com is in your OpenShell network allowlist — if it was added after initial NemoClaw setup, run nemoclaw policy reload after updating.
This is usually AI model latency, not an OpenClaw issue. Switch routine tasks to a faster model: set gpt-4o-mini or claude-haiku as default and reserve gpt-4o for complex queries. You can specify model per agent in the SOUL.md model: field.
Meta requires a public HTTPS URL for webhook verification. If you're running locally, use ngrok: ngrok http 18789. Copy the HTTPS URL into your Meta Developer webhook field. For production, point a domain to your VPS and set up SSL with Let's Encrypt (free).
Update your SOUL.md knowledge base with the correct information. The more specific and explicit your SOUL.md is, the more accurate your agent becomes. After editing, run openclaw agents reload maya to apply changes without restarting the Gateway.
Re-run openclaw onboard --install-daemon to re-register the system service. On Linux, check the service status with systemctl status openclaw and view crash logs with journalctl -u openclaw -n 50.
DM @Th3Alch3mist_ on X with your error message and the output of nemoclaw status + openclaw logs --tail 20. LocalClaw clients get priority support with same-day response.
What Comes Next
LocalClaw Managed Services
You've seen the full stack — from OpenClaw installation to a secured, multi-channel AI agent running for your business. The question now is: do you want to run this yourself, or have it done for you today?
Starter
$997 setup + $149/mo
One agent deployed on the channel of your choice. Full OpenClaw + NemoClaw setup. Monthly support, tuning, and monitoring included.
Business Engine
$1,997 setup + $299/mo
Multi-channel deployment across WhatsApp, Gmail & Telegram. Three specialized agents. Advanced OpenShell policy. Weekly performance reports.
Full Stack
$3,500 setup + $499/mo
Dedicated VPS, all channels, custom agents per department, NemoClaw enterprise policies, priority support, and monthly strategy calls.
Your agent is
ready to deploy.
Book a free 15-minute strategy call. We scope your deployment and you go live same day — no technical setup required on your end.
We schedule across all time zones. Most clients go live within 24 hours of their first call.