19 September, 2026

Building an AI agent with Claude: how it really works and the cost traps

Clients come to MONA asking about creating an AI agent with Claude after reading guides online. We see most people get stuck at connecting real data, or panic when the cloud computing bill arrives. Creating is easy. Keeping the bot stable every day is the hard part. To help you picture the workflow, we opened a public repository of 14 templates written in Vietnamese on September 19, 2026.

This article goes from building it yourself with Claude Code to the server cost trap, such as a MONA Cloud VPS priced from 550 VND/hour where shutting the machine down without deleting the service still charges for disk. Build it yourself only if you want to run an assistant that reads internal documents. Never wire the AI directly into a script that chats with customers to close orders, because the person carrying the risk of every wrong answer is your own wallet.

Building an AI agent with Claude: easy to make, hard to get real results

MONA bear compares a chatbot and an AI agent
An AI agent does more than chat: it also calls tools on its own to handle work.

Clients who come to MONA for software often assume a few lines of commands will give them an automatic assistant. Say you type a prompt asking for source code, and fifteen minutes later a bot appears on screen answering customers fluently. It looks impressive, until you realize this bot cannot look up a tracking number, cannot create an invoice by itself, and falls asleep the moment you close your computer. After weeks of tinkering, what you bring home is a smooth-talking bot that circles a few greetings. A real assistant AI agent has to pick up tools and go to work. A toy does not make money. MONA Agent template: free (MIT); MONA Cloud VPS running the agent: from 550 VND/hour, stop it and billing stops (we checked September 19, 2026).

How is an AI assistant different from an ordinary chatbot?

Social media is full of posts teaching how to create an agent with Claude, leading many people to think they can grab a template, write a job description in prose and have it run smoothly. We hit this mistake at the very first step ourselves: when we let the AI worker make up tool names in the AGENTS.md file, it immediately built completely wrong code. A real assistant must call real MCP tools accurately, such as an inventory-check API or one that creates a payment link. One slight error and the agent makes things up.

Why is building it yourself easy but running it for real hard?

We opened fourteen open-source Vietnamese templates to the public, and if you take a bot token and connect it to Telegram for a trial, you will find it extremely smooth. The hardest part comes when you attach the agent to a real business system for operations, especially when you want to put it AI chatbot on Zalo OA to close orders, which requires building an extra bridge of your own. When it runs for real, an agent working continuously needs a spending cap. Our MONA Cloud system returns error 402 as soon as the wallet runs out, forcing the agent to stop and notify a person. A production agent has to be strict.

Many people tinker with a few lines of commands and think they now have a great employee, but the machine's limits lie in the tangled connection infrastructure underneath. If you need a bot that reads internal documents and replies via Telegram, a self-built template is enough, and all data stays on your own machine. For a scenario where an agent revisits old orders to close sales through Zalo, an amateur-built flow breaks, and when a customer messages and nobody replies for three minutes, you lose the customer. The system has to be solid.

Keeping an AI agent running continuously and the fear of the end-of-month bill

A shop owner checking the API bill with MONA bear guidance
Setting a spending cap is a mandatory step to keep an AI agent from blowing the budget.

Looking at an API bill of ten-odd million VND at the end of the month with no revenue in sight is a scene we run into constantly when taking over systems that clients built themselves. When we dug through the logs, we saw the machine left on, with the AI calling commands day and night scraping junk data, and money draining away steadily just because one small computation task repeated tens of thousands of times. Without brakes, your wallet burns. MONA Agent template: free (MIT); MONA Cloud VPS running the agent: from 550 VND/hour, stop it and billing stops (we checked September 19, 2026).

After you create an AI agent with Claude, you need a server so it can wait on messages or orders continuously. According to measurements from the MONA Cloud review page, network latency from a server in Vietnam to VNPT users in Ho Chi Minh City is only 5-15 ms, while a Singapore server takes 30-50 ms. Choosing the right cloud VPS in-country lets the AI respond to customers instantly, but the billing trap lies in the starting rental rate of 550 VND per hour. We once helped a client who thought turning off the machine stopped the charges: they pressed the stop-server button and went to sleep without knowing the disk and static IP were still being held by the system and billed steadily at 15 VND per GB per hour. You must delete the service entirely to stop paying.

The rule for running an agent 24/7 is to always set a budget ceiling. On real infrastructure, your MONA Cloud wallet returns a 402 error the moment the balance runs out, forcing the agent to stop at once and alerting the admin immediately, so a bill can't balloon overnight. If you need to deploy AI agent platform into your sales flow, just share what you need and we'll build the whole closed-loop workflow for you. If it runs, it has to be able to brake.

Where does business data live when you hand work to AI?

The data security flow when using an AI agent
Understand how your data flows so you can protect your customer list when deploying AI.

Dropping a whole financial report or customer list into a public chat window gives you the chills. Plenty of you would rather stay up late typing every row into Excel by hand than hand your business numbers to some unknown system out there, for fear a competitor gets hold of them. That fear is justified.

We draw a clear line from the start. Your data sits on your own server by default, but the promise that not a single byte leaves your machine only holds for the Closed plan running an Ollama model. When you create an AI agent with Claude, the system runs on the API key you enter yourself, all your business reports stay on your server, and only the commands that are needed are encrypted and sent out.

Look at how we work through the MONA Agent template repository, which we opened to the public on September 19, 2026. We provide 14 templates written in Vietnamese, each with an AGENTS.md file so the AI can read and understand it, and a tools.json that lists only MCP tools that actually exist. Customers pull a template and build with Claude Code, and we measure real usage through clone counts and the number of VPS generated automatically from the attached deploy.md file. The whole deployment takes just 30 minutes, and your data stays entirely on your own machine.

Struggling to connect systems when following templates found online

Connect an AI agent to Zalo and your internal software
An AI agent is only worth something when it plugs straight into the workflow you already have.

A sales manager eagerly assembles a smart assistant from open source code she found online, yet the next morning ten staff are still sweating over copying each order number from the chat window into the accounting software. Clever, but useless. MONA Agent templates: free (MIT); MONA Cloud VPS running the agent: from 550 VND/hour, and it stops billing when you turn it off (we checked on September 19, 2026).

When we take over and fix DIY projects, the most painful break is clearly in the data pipeline, not in the choice of LLM. You download a sample config and it runs smoothly on your machine, but once it's plugged into the company's real workflow, the sales team goes back to being typists. The workflow doesn't lose a single step, it just gets messier, because you're now paying a salary to feed a human-powered data relay station.

When we packaged the MONA Agent template repository for people who want to create an AI agent with Claude, we made a serious mistake by letting the AI coder name the tools itself in the AGENTS.md file, and the code flow was built wrong from the very first step. The lesson: a standard template may only list MCP calls that have actually run. Take Telegram as an example: paste in the bot token string and the bot reads messages right away. But if you need a AI chatbot on Zalo OA today, don't use generic templates from the internet. This platform requires a dedicated bridge rather than offering a ready-made function, so until we've finished it, we absolutely don't make promises to customers.

Patching open source code together without a compatible communication gateway just forces your staff to copy by hand every day. It's better to put in the effort and build it properly once.

Take a look at your API bill: are you paying ten-odd million VND just for staff to copy chats by hand, or has the AI already been closing orders on its own? Your machine stays safe. Just pull the open source templates we keep in the MONA Agent repository and run them directly, so there's no fear of leaking data.

Take a MONA Agent template and build it in 30 minutes Call 1900 636 648

How our AI Bear team staffs the website and consultations

The list of 14 open source MONA Agent templates
The list of 14 AI assistant templates, packaged from projects we run ourselves.

A customer messages to buy at eleven at night while your staff are asleep, and when you open your computer the next morning, they've already ordered elsewhere. A lost deal, just like that. Instead of theorizing, we put a whole team of AI Agent on chat duty for MONA's own internal channels and a number of partner systems. Figures from an IBM report show that virtual assistants handle 80% of frequently asked questions cleanly, directly cutting 30% of the customer care payroll. With the systems we run, the results are measured straight from the actual server bill.

Fourteen MONA Agent templates, each solving one task

When you build an assistant with Claude on our system, you never have to start from scratch, because the monagent.vn homepage already shows its beta status, sharing one MONA Pass account and one Vietnamese dong wallet. We have opened 14 ready-made assistant templates to the public, such as cskh-zalo, sales-chot-don, ke-toan-hddt and phap-che-hop-dong, with all source code pushed to GitHub under the MIT license. Every folder ships with a README so readers understand it, an AGENTS.md that assigns the work to the AI, a SOUL.md that sets its personality, and a tools.json that connects directly to the real environment. The platform also integrates the OpenClaw core and the monacloud-mcp library to speed up responses.

We note every current technical limit clearly: the Zalo OA template is stuck waiting for a bridge, and the e-invoicing feature is waiting for the tax authority to open its programming interface. You pay nothing for the templates, and the system only charges for the hours your MONA Cloud VPS is switched on, from as little as 550 VND per hour. Turn it off and billing stops. Enter your own model API key, or pull the Ollama stack and run it directly on your internal server for safety. If you get stuck somewhere that needs deeper tailoring, call 1900 636 648 and we'll look at your system and work out the numbers.

An AI assistant that reads internal documents and messages you on Telegram

In practice it takes you less than 30 minutes to set up this whole system. Pull the repository to your machine, open the command panel Claude Code and ask it to read the AGENTS.md file to assemble the features, then run the deploy.md file, and the cloud_vps_create function pushes everything up to the cloud. A few seconds later the bot is sitting in your internal Telegram group, waiting for commands. If you mistype a command during setup and hit a black-screen error, just call MONA tech support and we'll walk you through the fix.

Falling into the server cost trap by assuming off means billing stops

MONA Cloud pricing and storage fees when a server is stopped
VPS costs are billed by the hour and are very low, but you need to delete the service entirely to stop being charged.

Many of you renting a server for an AI system for the first time just hit stop when you're not using it. You assume that powering down means the money stops flowing. The next morning you open the usage stats and the account is still quietly draining, bit by bit. You lose money without knowing why.

On September 19, 2026, we opened the MONA Agent template repository with 14 Vietnamese configuration sets in the AGENTS.md file. When we reviewed the VPS creation history from the deploy.md file, we immediately spotted this money-burning habit among bosses who tinker with Claude Code, painstakingly loading every MCP tool from cloud_*, mail_* to pay_link into tools.json. They start the bot, try a few messages, see it run smoothly, happily hit stop, and go to bed with peace of mind.

Server infrastructure doesn't behave like an appliance at home, where pulling the plug stops the cost. On the Cloud VPSplatform, the stop action only cuts the CPU fee of 550 VND/hour, while the system must keep the disk that holds your data and quietly bills an extra 15 VND/GB/hour.

To stop paying, you have to delete the service.

Never just stop the server and forget about it. If your company has no technical staff on hand to operate infrastructure, don't take it all on yourself. Deleting the service outright is the only way.

Automatic document processing also applies to legal and accounting

An AI agent that handles accounting and legal documents
Beyond customer care, an AI assistant also frees other departments from manual work.

Every time they reconcile invoices or check contracts, accounting and legal staff have to review every figure across dozens of different PDF files. It wears away hundreds of hours for nothing, and you're still anxious about paying compensation because a human eye misread one zero. Customers who come to MONA for software often use this same automated flow to screen job applications, answer customer service questions, or look up logistics shipments.

On September 19, 2026, we opened a repository of 14 Vietnamese templates to the public for non-technical users, including ready-made standard templates for accountants processing invoices and legal staff reviewing contracts. When we take over a system, we see that owners often fear having to tear down their old process that runs fine. In fact it takes exactly 30 minutes to set up a AI agent that reads your internal documents on its own. Your data stays on your own server, with no leak risk. Quick and clean.

This reading flow uses the same MCP toolset we normally set up, so the machine accurately recognizes every stamp and never makes up fake numbers to mislead the approver. We ourselves have been running Gấu Quản Trị to read internal documents since 2026: ask the company a question and it looks the answer up and replies with exact figures. Customers who come to MONA for systems now use a similar flow so the machine extracts everything in an instant. The data matches to the last dong.

Make good use of the AI agent platform open source version to test the document reading flow first, and don't rush to spend hundreds of millions of VND on a closed system before you know its limits. If your company expects a machine to decide everything and cut out the final human reviewer, don't use it at all. The cost of compensating for mistakes will be very high.

Common questions when building your own AI assistant with Claude Code

Answers to questions about creating an AI agent
We keep the real status of our tools up to date so you have the most accurate view.

Hundreds of you go online looking for how to create an AI agent on Gemini or Claude, expecting to download, unzip, and have a virtual employee running at full speed. But when you open the llms.txt file in the September 19, 2026 test version of the MONA Agent project, you'll see we state plainly that the one-click install button isn't usable yet, and creating an AI agent with Claude Code is much harder. Seriously tough.

Whenever we take on a DIY project, customers who come to MONA ask where the company data is stored and who they actually have to pay when using open source. The promise that your data sits on your own server holds for every plan, but the guarantee that not a single byte leaks outside the system is only possible with the Closed plan running Ollama. On cost, the templates themselves are free, but you pay for server rental from 550 VND an hour plus the fees for calling large language models. Many of you think turning off the server settles it. In reality the cloud system still holds the disk and the network address, and your account keeps being billed 15 VND per gigabyte per hour. To stop paying, you must wipe the service completely.

Another burning question is how to connect this assistant to your internal messaging system. With Telegram, you just attach a programming key string and it connects instantly, but for Zalo OA you have to build a separate bridge yourself. We advise that if your tech team hasn't finished the Zalo bridge, don't promise anything to the sales department yet. Creating an AI agent with Claude is far thornier than the success stories online make it sound.

Don't let a virtual assistant handle anything involving money automatically without a person reviewing it. If you need a AI agent platform that automatically goes into your customers' orders or checks inventory, don't use this free DIY version. A purpose-built, closed enterprise software system is what we recommend to keep your business information absolutely safe.

How many hours of manual work could your team save this week?

AI agent consulting for businesses
Start automating your workflows today so your staff can focus on generating revenue.

Eleven o'clock on a Friday night, the person on page duty has just dozed off when three customers message asking to change their service details. The next morning, two have quietly left and one has ordered from a competitor. The money drops right in front of you. Every business owner has felt that pang, watching an opportunity slip away just because you can't make your staff stay up all night typing. Paying for a third shift just to watch trivial messages is a brutal drain on your budget, and it ties your best salespeople to repetitive work instead of sending them out to win big contracts.

We solve this problem for good with an AI agent system that automates the work, running in production since January 2026 to carry the full load of repetitive inbox traffic. Specifically, we coded the flows ourselves for Gấu Cười for customer care, Gấu Webmaster keeping watch over the homepage, and Gấu Thức Khuya, which automatically reviews the list for sending reminder emails to past customers. From building, digging through the source code and debugging our own internal systems, we packaged 14 complete Vietnamese MONA Agent templates. Shop owners don't need to be technical: just select the features in the right order, and within five minutes the system is running.

When you plug new technology into a running process, the payment gateway API connection is always where the risk blows up most easily. In mid-September 2026, we had a webhook key mismatch when connecting two internal pieces of software: the customer's money went straight into the bank account, but the balance in the system refused to move. We had to dig through every line of the log returned from the bank to patch the sync error right away. We learned the hard way. As a result, our build a mini piece of software themselves process now requires a real card swipe, and only when the transaction shows as successful in the banking app do we dare hand it over to you.

If you want your sales team to focus on closing contracts worth tens of millions of VND instead of wearing themselves out answering trivial messages, put this team of digital assistants into operation now. But there's one thing to avoid. If your company's sales process is still running on the gut feeling of a few star salespeople and pricing rules change every day, don't bring an agent in to wreck the system.

A manifesto for growing together

MONA is a software, digital transformation and AI transformation company, but above all, MONA is a B2B service company that benefits directly from this: THE MORE SUCCESSFUL VIETNAMESE BUSINESSES BECOME, THE MORE MONA BENEFITS, while a declining, slow market with weak technology is exactly what kills MONA's future business opportunities. So, more than anyone, MONA wants you to succeed, and MONA actually steps in to help you succeed.

And AI transformation is the key to that success for all of us.

MONA

At the end of the month you're holding a ten-million-plus VND API bill while revenue hasn't changed, and your staff are still copying data by hand from one system to another. We know you've tried hard to put technology together to automate your processes. Plenty of customers came to MONA after dipping into technology just because it was the trend. Their systems broke in the very first month, because the software was built for casual chat, couldn't close the sales loop, and also faced the risk of leaking confidential data.

We untangle this with the Business plan, deploying an assistant flow from the MONA Agent repository directly on a domestic server from 550 VND an hour, keeping all your information in your hands. Machines have to make money. Call 1900 636 648, tell us which workflow is stuck, and we'll work out how to fix it right away.

Take a MONA Agent template and build it in 30 minutes Call 1900 636 648

MONA has a podcast: listen instead of readingEpisodes on AI, automation & SEO for Vietnamese businesses
Nghe ngay
About the author

CEO & Founder · The MONA

Founder of The MONA, who built a services company of 60+ staff with 14,000+ projects delivered. Khánh Hùng shares how he thinks about building systems that run themselves, automating a business, and the real lessons from delivering projects at MONA.

Views 379
Rate this article 4,8/5 · 22 ratings

Professional website
design services

Own a website with a beautiful, 100% exclusive design and a lifetime warranty, with the ability
to expand features flexibly as your business grows, starting today!

Contact Mona

    MONA có riêng một Người "Bạn Thân" cho bạn - Người Account sẽ đồng hành, hỗ trợ, hướng dẫn, đặt đồ ăn cho bạn mãi mãi, từ đây về sau!
    MONA will never sell your information or send you spam
    The MONA team advising on an online education management system
    Hỏi đáp giáo dục 4.0
    Tạo cuộc hẹn miễn phí với MONA để giải đáp và tư vấn mọi thắc mắc về giải pháp số hoá ngành giáo dục
    Thời lượng cuộc hẹn
    45 Phút
    Ngày và giờ
    Thứ 2, ngày 25 tháng 12, 2023
    [9:30 - 10:15]

      Chọn ngày và giờ
      Khung giờ
      Back
      Hãy cho MONA biết bạn là ai
      0:00