OpenClaw: How to Set Up an AI Assistant That Responds to Clients 24/7

A client sends a message on WhatsApp at 11 PM asking about pricing. A potential partner reaches out on Telegram at 7 AM before your team logs in. Someone in your Slack support channel posts a question on a Sunday. For small business owners and solo operators, these aren't edge cases — they're regular occurrences that either get answered late or don't get answered at all. Both outcomes cost something.

OpenClaw is an open-source personal AI assistant designed to close that gap. It doesn't sit on your website as a floating chat bubble; it lives on a server or machine you control, connects to the messaging apps your clients already use, and stays online around the clock. This article covers what OpenClaw actually is, how to get it running, and how business owners are using it to handle communication that used to wait until Monday morning.

What OpenClaw Is — and What It Isn't

OpenClaw describes itself as "the AI that actually does things." That framing matters, because it positions the product differently from both traditional chatbots and the new wave of AI assistants baked into operating systems. It's not a chatbot widget you embed in a storefront, and it's not a cloud service where your conversations pass through someone else's infrastructure. It's software you install, configure, and run — on your own terms.

The core idea is simple: you have a Gateway (the engine that manages everything), and that Gateway connects to the channels you already use — WhatsApp, Telegram, Slack, Discord, Signal, iMessage, and more. When someone messages you on any of those platforms, OpenClaw receives the message, processes it using whichever AI model you've configured, and replies. You can have the assistant handle conversations entirely on its own, or set it up to notify you when something requires your attention.

Created by Peter Steinberger and an active open-source community, OpenClaw has grown to over 376,000 GitHub stars since its release. In early 2026, TechCrunch noted that the creator joined OpenAI, while the project itself continued as independent open-source software — a detail that tells you something about how seriously the project is taken.

The Architecture: Gateway, Channels, Skills

Understanding three concepts makes everything else about OpenClaw click into place.

The Gateway

The Gateway is OpenClaw's local control plane. It runs as a background daemon on your machine or server, listens on port 18789, and orchestrates everything: sessions, channels, tools, and model calls. When you start OpenClaw, you're starting the Gateway. When it goes offline, your assistant goes offline. This is why hosting it on a server that runs continuously matters — more on that later.

The Gateway handles multi-agent routing, which means you can have different assistants (workspaces) handling different channels. A Slack workspace for internal team communication can route to a different agent than your WhatsApp number used for clients.

Channels

Channels are the surfaces where conversations happen. OpenClaw ships with built-in support for Discord, Google Chat, iMessage, IRC, Signal, Slack, Telegram, WebChat, and WhatsApp. Additional channels — Matrix, Microsoft Teams, Mattermost, Nextcloud Talk, Feishu, Twitch, and others — come as bundled plugins. Each channel connects to the same Gateway, so one running instance handles all of them simultaneously.

For business use, WhatsApp and Telegram are the fastest channels to get working. Telegram requires just a bot token from BotFather. WhatsApp takes slightly more setup but reaches the widest client base. Slack integration is particularly relevant for teams managing a support channel — the Showcase section of the official documentation includes a real example of a company using OpenClaw to watch a Slack channel and respond autonomously, forwarding alerts to Telegram when something needs a human eye.

Skills and Memory

Skills are extensions that give OpenClaw specific capabilities. Some come from ClawHub (the official skills registry), others you can build yourself — or have OpenClaw build for you by describing what you need in a chat. Community members have built skills for Gmail, Todoist, Linear, Home Assistant, Jira, public transit APIs, and dozens of other tools. The assistant can install and start using a new skill within the same conversation where you requested it.

Persistent memory is a distinct feature from skills. OpenClaw builds up a knowledge of your context over time — your preferences, the context of ongoing client relationships, recurring questions it has learned to handle. That accumulated memory makes the assistant progressively more useful rather than starting from scratch every session.

What OpenClaw Can Do for a Business Owner

The product is general-purpose by design, which means its value depends on how it's configured. For a business owner focused on client communication, a few categories stand out based on what the community is actually building.

On the messaging side: receiving and replying to client messages on WhatsApp or Telegram, answering recurring questions about services or availability, routing urgent inquiries to you via a notification on another channel, and acknowledging receipt of messages that you'll follow up on personally. The assistant can be given a persona and instructions that match your brand's communication tone.

On the operational side: email inbox management (reading, summarizing, drafting replies), calendar coordination, web research for pricing or competitor information, creating and tracking tasks in tools like Todoist or Linear, and running scheduled checks on information you care about. Community users have built workflows that process email attachments, book appointments via browser automation, and generate morning briefings before the workday starts.

Browser automation is worth mentioning separately. OpenClaw can navigate websites, fill forms, and extract information without APIs — which means it can interact with tools that don't offer integrations, as long as they have a web interface.

Key Features: What's Included

Setting It Up: From Zero to Running Assistant

OpenClaw runs on macOS, Linux, and Windows (via WSL2). The documentation recommends Node.js 24, though Node 22.19 and above also works. The full setup takes about five minutes for the installation itself; connecting channels and tuning the assistant takes longer depending on how deeply you want to configure things.

Installation

On macOS or Linux, the quickest path is the one-line installer. It handles Node.js and all dependencies automatically:

On Windows via PowerShell:

Alternatively, if you prefer to manage it through npm or pnpm directly:

Onboarding

After installation, run the onboarding wizard. The --install-daemon flag sets up the Gateway as a background service so it survives reboots and keeps running when you close the terminal:

The wizard prompts you to choose a model provider and supply an API key. If you already use Claude or ChatGPT, you'll use the same API key here. After onboarding completes, verify the Gateway is running:

If it responds with the Gateway listening on port 18789, the engine is live. Open the Control UI dashboard with openclaw dashboard and send a test message to confirm everything works end-to-end before moving on to channel setup.

Connecting a Channel

Telegram is the fastest channel to connect — it requires only a bot token from Telegram's BotFather. The documentation at docs.openclaw.ai/channels/telegram walks through the exact steps. WhatsApp, Slack, and Discord each have their own setup guides under the Channels section. Once a channel is connected, messages arriving on that platform route through the Gateway and the assistant begins responding.

DM access is locked down by default: unknown senders receive a pairing code and the assistant doesn't process their message until you explicitly approve them. For a business setting, you can open this up by setting dmPolicy="open" on channels where you want the assistant to respond to all inbound messages without prior pairing.

Cloud Servers from $5 / moIntel Xeon Gold 6254 3.1 GHz CPU, SLA 99,9%, 100 Mbps channelTry

Five Use Cases Where OpenClaw Makes a Real Difference

1. 24/7 Client Messaging on WhatsApp or Telegram

This is the most direct use case for business owners. Configure OpenClaw with a persona that reflects your brand's communication style — name, tone, the information it should always have available — and connect it to your WhatsApp Business or Telegram number. When a client messages outside business hours, the assistant responds immediately rather than letting the inquiry sit until morning.

The memory system means it gets contextually better over time. After several conversations, it knows which clients have ongoing projects, what their typical questions are, and what answers you've given before. That context shapes responses in a way that generic chatbots can't replicate.

2. Slack Support Channel Automation

The Showcase documentation includes a verified community example: an OpenClaw instance watching a company Slack support channel, responding helpfully to incoming questions, and forwarding alerts to Telegram when something needs a human decision. That same instance autonomously fixed a production bug in a deployed app without being asked — it noticed the issue during the channel monitoring and resolved it.

For a business running internal or client-facing Slack channels, this pattern works well: the assistant handles routine questions and acknowledgments, flags anything requiring a judgment call, and keeps the channel responsive without manual attention.

3. Email Inbox Triage

With a Gmail skill (available on ClawHub) or browser-based email access, OpenClaw can read incoming messages, categorize them by urgency or topic, draft replies, unsubscribe from lists, and surface the items that actually need your attention. Community users report running this on schedules — a morning email sweep that delivers a summary via Telegram before the workday starts, so you begin with context rather than an unprocessed queue.

4. Scheduled Follow-Ups and Proactive Outreach

OpenClaw's cron job functionality handles time-based automation without any additional tools. You can configure it to send a follow-up message to a client three days after an initial conversation, deliver a weekly summary of open items, or check in on a project status at a defined interval. These aren't blast messages — they're context-aware follow-ups shaped by what the assistant already knows about the specific client or project.

The heartbeat feature is a related capability: the assistant sends you a proactive check-in at defined intervals, which some users configure as a daily briefing covering key updates, pending items, and anything the assistant flags as needing attention.

5. On-Demand Web Research and Information Lookups

When a client asks a question you don't have the answer to immediately — a delivery carrier's current estimated timeline, a competitor's public pricing, availability of a specific product from a supplier — OpenClaw can look it up in real time using its browser control capability. No API integration with the target site is required. It navigates to the page, extracts the relevant information, and returns an answer within the same conversation.

This turns the assistant from a scripted responder into something more capable of handling genuinely varied questions, which is where a lot of scripted bots fail.

Limitations and Risks to Factor In

OpenClaw is a general-purpose assistant, not a product built specifically for any one industry or workflow. That flexibility is its strength, but it also means there's no out-of-the-box configuration for business communication — you're shaping it yourself. The quality of the assistant's responses reflects the quality of the instructions and context you give it. An undertrained, poorly prompted assistant will give generic, sometimes wrong answers.

The security model requires attention before you open the assistant to inbound messages from unknown senders. DM access is locked by default, and for good reason — OpenClaw connects to real messaging surfaces where it has real capabilities (file access, browser control, command execution). The documentation has a full security guide and an exposure runbook that should be read before making the assistant publicly accessible. Run openclaw doctor after setup to surface any misconfigured DM policies or risky settings.

Model cost is a variable that's easy to underestimate. OpenClaw doesn't charge per message, but it does call the AI model provider for every substantive interaction. With a busy client-facing channel, that means a steady stream of API calls. Monitor usage, set reasonable rate limits, and choose a model tier that fits your expected volume. For high-frequency, lower-complexity queries, a faster and cheaper model may serve better than a flagship one.

Finally, the assistant has no human empathy and no institutional awareness of your business beyond what you tell it. Complex client situations — disputes, sensitive negotiations, anything requiring nuanced judgment — need a human. Build clear escalation paths so the assistant knows when to notify you rather than respond on its own.

Running OpenClaw 24/7: Infrastructure Considerations

The most common reason an OpenClaw deployment fails in practice isn't configuration — it's uptime. Running the Gateway on a personal laptop or desktop means the assistant goes offline whenever the machine sleeps, restarts, or loses power. For a 24/7 client communication tool, that's not acceptable.

The practical solution is a always-on Linux VPS. OpenClaw runs comfortably on a standard setup: a lightweight Linux instance with a recent Node.js runtime is sufficient for most single-business deployments. The Gateway process is not particularly resource-heavy at typical messaging volumes; the resource usage scales with the number of concurrent active conversations and any browser automation tasks you run.

Serverspace VPS provides Linux-based virtual servers with predictable monthly pricing — a sensible match for OpenClaw's fixed-cost model, where you pay for the server regardless of how many messages the assistant handles. With data center locations across different regions, you can place the server close to your client base to minimize any latency in response delivery. Setting up the Gateway as a systemd service (which the --install-daemon flag handles automatically on Linux) ensures the assistant restarts automatically if the process crashes or the server reboots.

On the operational side, configure monitoring for the Gateway process. A basic health check that pings localhost:18789 and alerts you if it stops responding takes minutes to set up and prevents the scenario where the assistant silently goes offline with no indication. Run openclaw gateway status regularly during the first weeks to get a feel for how stable your particular setup is.

Cloud Servers from $5 / moIntel Xeon Gold 6254 3.1 GHz CPU, SLA 99,9%, 100 Mbps channelTry

Common Setup Mistakes

Skipping the Persona and Instructions

OpenClaw out of the box will respond as a general assistant — helpful but generic. For business use, it needs explicit context: who you are, what you do, what the assistant is allowed to answer on your behalf, what it should always escalate, and what tone to use. The SOUL.md file in the workspace root is where you define the assistant's persona. Spending an hour on this has more impact on the quality of responses than any model configuration choice.

Opening DM Access Before Reading the Security Guide

The default dmPolicy="pairing" setting exists for a reason. Before switching to open DMs, read the Security documentation and the exposure runbook. The assistant has real capabilities on the host machine; an improperly secured instance is a meaningful risk, not a theoretical one.

Using a High-Cost Model for Everything

Flagship models from Anthropic or OpenAI are capable but expensive at scale. For a client-facing channel handling hundreds of messages per day, a smaller, faster model will handle the majority of questions adequately and at a fraction of the cost. Configure model fallbacks and consider routing different task types to models appropriate for their complexity.

Not Building Escalation Logic

The assistant needs a defined path for situations it can't or shouldn't handle. Without explicit escalation instructions, it will attempt to respond to everything — sometimes inappropriately. Tell it exactly which scenarios should trigger a notification to you rather than an autonomous response: client complaints, payment-related questions, requests that involve commitments you haven't pre-approved.

Treating Setup as a One-Time Task

OpenClaw improves with maintenance. Skills need updates. The model landscape evolves. Your business context changes. The assistant's instructions should be revisited every few weeks, particularly in the early months. The Changelog on GitHub and the OpenClaw blog at openclaw.ai/blog track changes worth staying current on.

Who OpenClaw Is Actually For

OpenClaw fits best when a few things are true. You're comfortable installing software from a command line, or you're willing to learn — the setup isn't complicated, but it's not a point-and-click experience. You want control over your data and prefer not to route client conversations through a third-party SaaS platform. And you're running a communication workload that genuinely can't be covered by business hours alone.

It's less suited to teams that need a polished, no-configuration product with a support team behind it. The open-source community is active and the Discord server is responsive, but support is peer-based rather than contractual. For a technically capable solo operator or small team, that trade-off is straightforward. For a larger organization with strict procurement requirements, it requires more evaluation.

On the cost side, the economics are straightforward: you pay for server hosting (fixed) and API calls to your model provider (variable but predictable). There's no per-seat licensing, no usage-based markup, and no feature gating based on a subscription tier. For businesses that handle high message volumes, that model becomes meaningfully cheaper than SaaS alternatives at scale.

Conclusion

OpenClaw is a self-hosted personal AI assistant that connects to the messaging apps your clients already use and stays online continuously. It handles inbound messages, manages email, runs scheduled tasks, and can extend its own capabilities with skills from the community or ones you build yourself. The data stays on your server, the cost model is predictable, and the setup is more accessible than the open-source label might suggest.

For a business owner losing opportunities to slow response times outside office hours, it's a practical answer to a structural problem — not because it does one specific thing perfectly, but because it can be shaped to cover the specific gaps in your workflow. Start with one channel and one clear use case. Get that working well, then expand.

FAQ