An AI customer service agent that lives on your own hardware
If your customers reach you on WhatsApp, Telegram, or WeChat, you already know the shape of the problem: messages arrive at 2am, the same six questions repeat all day, and the only way to answer faster is to hire more people. A support agent that runs on your own box changes the arithmetic — it answers on the channel your customers already use, and the conversation records stay on your premises.
What actually costs you money today
| Approach | Monthly | Where it breaks |
|---|---|---|
| You answer personally | $0 | Does not scale; you are the bottleneck at night and on weekends |
| Hire front-line staff | Salary per head | Hiring cycle, training, turnover; coverage gaps remain |
| Outsource to a call centre | Per seat or per ticket | Scripted answers only; your context never transfers |
| SaaS chatbot platform | Per seat + per conversation | Your customer conversations sit on someone else's servers |
| Agent on your own box | Credits only | Needs setup and a knowledge base; not a replacement for humans on hard cases |
How it works
- Plug in the box — power and network. It boots into the agent runtime.
- Connect your channel — WhatsApp, Telegram, or WeChat. Scan a code and pair it.
- Point it at your documents — product info, policies, price lists, FAQs. This is what makes the answers yours rather than generic.
- Set the escalation rule — when confidence is low or the customer asks for a human, hand off. Decide this before you go live, not after.
What it does well, and what it does not
Being clear about this up front saves everyone a bad deployment.
| Good fit | Not a good fit |
|---|---|
| Answering the same questions repeatedly, accurately, 24/7 | Negotiating price or handling an angry escalation |
| Qualifying leads and booking appointments | Anything needing a judgement call you have not written down |
| Looking up order status from your own systems | Replacing a human relationship on high-value accounts |
| Drafting replies for a human to approve | Autonomous commitments you have not authorised |
The realistic pattern is that the agent absorbs the repetitive traffic and escalates the rest, so a small team handles more volume without the quality dropping.
Why local matters here specifically
Customer conversations are the most sensitive data a small business holds. They contain names, order details, complaints, sometimes payment references. On a SaaS chatbot platform, all of that flows through a third party's infrastructure and is subject to their retention policy.
On a local box, the conversation records, the memory, and the documents the agent reads stay on hardware you own. Only the model call itself goes out. If a customer asks where their data lives, you can answer concretely.
Cost of running it
The device draws under 10W, so power is a rounding error. The real recurring cost is credits, debited by the tokens the agent spends answering. Two things keep it predictable:
- Route simple questions to a cheap model. Most support traffic is classification and lookup, which does not need a frontier model.
- Cache the answers that never change. Shipping policy, return window, opening hours — answer from your own documents rather than re-deriving them every time.
Frequently asked
Which channels are supported?
WeChat, WeCom, WhatsApp, and Telegram. The channel is just the transport — the agent is the same everywhere.
Can it hand off to a person?
Yes, and you configure the trigger. Low confidence, an explicit request for a human, or a keyword you define. The conversation history goes with the handoff so nobody starts over.
Do I need to write code?
No. Setup is pairing a channel and pointing the agent at your documents. Nothing beyond that is required.
What happens when the internet drops?
The agent runtime and your stored conversations keep working locally. Model inference pauses until connectivity returns, so drafting and retrieval degrade but nothing is lost.
Can several people use it at once?
Yes. Contacts and history are stored on the device, and the enterprise edition covers higher concurrency with private deployment and an SLA.
Put a support agent on your own hardware
Luz AIbox Pro — connect WhatsApp, Telegram or WeChat, point it at your documents, and let it take the repetitive traffic.
View Pricing → Deploy for your team →