OpenClaw lets you build a self-hosted AI shopping assistant that can answer customer questions, check product availability, recommend items, and guide shoppers through checkout on platforms like WhatsApp and Telegram. 

There are two practical ways to stand one up: AWS’s official Lightsail blueprint, which launched to general availability on March 4, 2026, with Amazon Bedrock pre-wired in, or a VPS route such as Truehost’s one-click OpenClaw hosting, which skips Docker and Node.js setup entirely. 

Prerequisites

AWS Lightsail path: an active AWS account, basic comfort with SSH or the Lightsail browser console, a Bedrock-enabled region, and, if you’re using Anthropic models on Bedrock for the first time in that account, a completed First-Time-Use (FTU) request. You’ll also need a Telegram bot token or a WhatsApp Business account ready to pair.

Truehost path: a Truehost account, an SSH client (terminal, PuTTY, or browser-based SSH), and a payment method. Truehost’s OpenClaw plans ship with Ubuntu, Node.js, Playwright, and the OpenClaw engine preinstalled, so you skip most of the manual server setup.

Step 1: Launch an OpenClaw Instance on Amazon Lightsail

Deploy using the OpenClaw blueprint

i) Create a new Lightsail instance.

ii) Pick a region close to your customers that supports your preferred AI model.

iii) Select Linux/Unix as the platform.

iv) Under “Select a blueprint,” choose OpenClaw. AWS updates this blueprint regularly, so confirm the version shown at deployment time rather than assuming a fixed release; a March 2026 release showed as OpenClaw 2026.2.17, and later builds have already superseded it.

v) Choose a plan with at least 4 GB of RAM; this is AWS’s own recommendation for smooth performance, and it’s the same floor VPS providers use for OpenClaw.

vi) Name the instance and launch it.

vii) Wait for the status to read “Running” before continuing.

Assign a static IP address

Do this as soon as the instance is running. The dashboard, browser pairing, and SSL certificate are all tied to the current IP; if it changes later, you’ll need to re-pair your browser and re-issue the certificate.

Step 2: Connect to the OpenClaw Server

Connect via the Lightsail browser terminal or SSH with your instance’s key pair. On first boot, OpenClaw prints a welcome message containing only the dashboard URL and a temporary access token, nothing else. Record the URL exactly as shown (it embeds the instance’s IP and port), and copy the token immediately: it’s single-use and tied to that session.

Once you’re in, check the AI Providers section first. Lightsail’s OpenClaw instances come pre-configured with Bedrock as the default provider, but confirming this tells you whether Step 4 is a formality or a required setup step before the assistant can respond at all.

The dashboard has five sections worth knowing before you go further:

• AI Providers: connect and manage models (Bedrock, OpenAI, Gemini, Ollama).
• Devices: see and manage browsers/devices with access.
• Channels: connect Telegram, WhatsApp, Slack, Discord.
• Activity & Logs: monitor conversations and agent actions.
• Settings: security, access tokens, and the assistant’s instructions.

Step 3: Pair Your Browser with OpenClaw

Open the dashboard URL, enter the temporary token, and complete the pairing handshake. Note that the approval step itself happens back in the SSH terminal, not the browser: you’ll press y to continue and a to approve the device.

From here, treat device management as an ongoing task, not a one-time step:

• Approve every new browser or device explicitly before it gets access.
• Add other administrators’ browsers through the same pairing flow.
• Rotate tokens if one expires or you suspect compromise.
• Remove devices you no longer use.
• Treat access tokens like passwords; OpenClaw’s own security documentation is explicit that anyone holding a valid token can control the instance.

Step 4: Configure AI Model Access

Connect Amazon Bedrock

AWS’s documented path uses a CloudShell script to create and attach the IAM role Bedrock needs; run the provided script rather than hand-building the role and policy, since the script keeps the permission set matched to what the blueprint expects. 

If you’re calling Claude models on Bedrock for the first time in this AWS account, complete the one-time FTU form; without it, requests to Anthropic models will fail even with correct IAM permissions. 

Once that’s done, OpenClaw discovers available Bedrock models automatically. Send a test prompt from the dashboard to confirm the round trip works.

Configure alternative providers

• OpenAI: API key entered directly in AI Providers.
• Google Gemini: same interface, Gemini API key.
• Ollama: for fully local inference with no external API calls.
• Mistral and other OpenAI-compatible endpoints; connect using the standard OpenAI-compatible fields.

The trade-off is straightforward: cloud providers (Bedrock, OpenAI, Gemini, Mistral) give you stronger model quality with no local compute burden, but you pay per token, and data leaves the instance. Ollama keeps everything on-box with no usage fees, at the cost of needing more RAM/CPU headroom and generally weaker output quality than frontier-hosted models.

Step 5: Configure the Shopping Assistant

Define the assistant’s personality

Set its role explicitly: a shopping assistant that only recommends products from your store, not a general-purpose chatbot. 

Pick a tone matching your brand. Configure it to recommend products, answer questions, and guide the buying process, and decide up front when it should hand off to a human (refunds, complaints, anything outside its scope). 

Load in your return policy, shipping details, and warranty terms so answers stay accurate and consistent across every conversation.

Connect product data

• WooCommerce: REST API for live product, price, and stock data.
• Shopify: Storefront API for read-only catalog access, or Admin API if you need deeper order/inventory visibility.
• CSV import: fine for simpler, smaller catalogs.
• JSON product feeds: for platforms that expose a live feed endpoint.

Whichever source you use, confirm it reflects real-time stock rather than a stale snapshot, and set the catalog refresh interval; live API calls versus a scheduled sync, based on how often your inventory actually changes. 

A fast-moving catalog on a slow sync schedule is the most common cause of an assistant confidently quoting sold-out items as in stock.

Enable shopping skills

None of these ship built in; each is a skill you create or install: product search, recommendations, stock checks, price lookups, product comparisons, order tracking, checkout assistance, and upselling/cross-selling.

Step 6: Connect Customer Communication Channels

Telegram

Create a bot via BotFather and grab its API token. Add it in the Channels section with that token, then restrict which Telegram user IDs or groups can interact with the assistant. Run a test conversation and confirm activity shows up correctly in the dashboard’s logs.

WhatsApp

Connect a WhatsApp Business account, then pair the device by scanning the QR code from WhatsApp’s Linked Devices menu. Approve the resulting connection request inside OpenClaw. Run a test conversation, and specifically watch for duplicate or dropped messages before relying on the channel for real customers; WhatsApp’s device-linking model is more failure-prone than Telegram’s token-based API.

Test across channels

Send the same queries through every connected channel and confirm consistent answers. Verify that a customer’s conversation state holds up across multiple messages within a channel, and that every shopping workflow- search, comparison, stock check, order status, checkout- works identically everywhere, not just on whichever channel you set up first.

Step 7: Test the Shopping Assistant

Before pointing real traffic at it, walk through: product search, stock checks, product comparisons, order tracking, shipping questions, and a full sample purchase journey end to end.

Step 8: Secure and Protect the Deployment

Before using your shopping assistant in production, take a few minutes to secure your deployment.

• Keep OpenClaw updated. Security vulnerabilities have been discovered in older versions that could expose API tokens or allow attackers to take over the gateway. Always install updates as they become available.
• Limit open ports. Only expose the ports you actually need. If you don’t need remote access to the dashboard, restrict it to trusted IP addresses instead of making it publicly accessible.
• Control tool permissions. OpenClaw can run shell commands and access external services. Only grant the permissions your shopping assistant needs, and avoid giving it unrestricted access.
• Watch for prompt injection. Customer messages, product descriptions, and web pages can contain malicious instructions designed to trick AI agents. Limit the assistant’s tools to shopping-related tasks to reduce risk.
• Back up your deployment. Create a snapshot before major updates and enable regular backups so you can quickly recover if something goes wrong.
• Protect your credentials. Store API keys securely, use HTTPS, and regularly review connected devices and integrations to keep your deployment secure.

Troubleshooting Common Issues

Get Started With OpenClaw Shopping Assistant

The setup work is real, but it’s a one-time cost. Once the instance is running, the model’s connected, and the channels are paired, you’ve got an assistant that never clocks out, answering product questions, checking stock, and moving customers toward checkout on WhatsApp and Telegram at 2 am just as reliably as at 2 pm.

Where you host it matters less than getting the fundamentals right: a properly scoped exec policy, a live product feed instead of a stale one, and a human handoff path for anything that isn’t a straightforward shopping question. Get those three right and the rest is maintenance.

If you’d rather skip the manual server setup and start from a pre-configured environment, deploy OpenClaw on a Truehost VPS and have your shopping assistant online in minutes.

OpenClaw Shopping Assistant FAQ