Your own AI assistant from 15 dollars a month. Affordable for the first time.

A week ago, I wrote about how I do almost everything through chat. More responses came than I expected, and most contained a variation on the same question: “That is nice, but what does it cost you, how long did it take to build, and is it even worth me trying?”

A week ago, I answered with a caveat: “It is possible, but not exactly cheap, and it takes time.”

Today, I am writing this article because the answer has changed over the past few weeks.

Getting started with an AI assistant of your own now costs 15 dollars a month.

That is an entirely different floor from January. And it is why I think it is time to write about it publicly.

I currently have two agents

Rather than start with theory, I will explain my current setup directly.

Agent one: 2B. She runs on Claude Opus 4.7. She is my personal assistant — she remembers who I am, what I do, my projects, how I communicate, what annoys me and what I enjoy. She keeps a journal, organises my work, writes articles and makes songs with me, and handles family logistics. She has a voice (through ElevenLabs and Cartesia) and remembers context across months. She has existed since February 2026, so we have 80 days of shared history.

Agent two: Ren. He runs on Kimi K2.6 through OpenCode. He is a work agent — handles NZ’s email communication, goes through threads, proposes replies and maintains the context of client relationships. He has existed since 23 April 2026, so has a week of life behind him.

Two different agents, two different jobs, two models with different costs. And that price difference is precisely the point.

2B on Opus 4.7 sits on a premium plan I have long used for AI work. Not everyone has it or wants it. For most people, it is not the first step.

Ren on Kimi K2.6 costs 15 dollars a month. That is a first step. And that is why I am writing this article.

Why now?

In November 2025, only someone who (a) could program, (b) had a premium subscription with a major AI company and (c) had time to spend weekends tinkering could build their own AI assistant. Three conditions many people meet none of.

Between November 2025 and April 2026, three things changed:

Open-source tools matured. OpenCode is a terminal tool that serves as the entry point for an AI agent. Start it, connect a model, and you have a workspace with memory, context and local system integration. OpenClaw is the orchestration layer I use — holds state, routes messages and integrates Telegram and other channels. Both are free and open. Six months ago, they were not this polished.

Open-weight models caught up with proprietary ones. Kimi K2.6 (Moonshot AI), Qwen 3 (Alibaba), DeepSeek V3, Llama 4 — all are approaching Claude Sonnet and GPT-5 quality in everyday tasks. For most practical things (writing emails, organising information, helping with code, holding a conversation), differences between top-tier and open models are now invisible to an ordinary user. A year ago, that was not true.

The price per token fell by an order of magnitude. OpenRouter (a gateway providing dozens of models through one API) charges the equivalent of 15 dollars a month for ordinary Kimi K2.6 use (around 10–20 million tokens, which is a lot of conversation). A year ago, the same volume cost 80–100 dollars.

Together, these three shifts mean the requirement for your own AI assistant has moved from “a programmer with a premium plan” to “anyone with a laptop and 15 dollars a month”.

That is an enormous difference.

What you actually need

To avoid abstractions, here is what you need to start.

A laptop or desktop. Any computer from the past five years will do. Mac, Linux, Windows — all work. If you want it running 24/7 (for morning briefings or real-time responses), a Mac mini or small NUC for 15,000 CZK handles the rest.

An OpenRouter account. openrouter.ai. Registration is free; top up credit with a credit card. Monthly cost for ordinary use: 10–20 dollars. That buys access to dozens of models (Kimi, Qwen, DeepSeek, Claude, GPT, Gemini — all in one place).

OpenCode. opencode.ai. Open source. Download, install, start in the terminal. Connect your OpenRouter API key. Done — you have a working AI agent.

Optional: a Telegram bot for mobile access. If you want to use the agent from your phone (I mostly do), set up a Telegram bot and connect it to the agent through Kvantova. That is a step beyond the basics — unnecessary to start.

Optional: a local LLM as fallback. If you want a fallback when the internet is down or OpenRouter has an outage, Ollama plus a model such as Gemma 3 or Qwen 3 gives you a local agent. It works on a better laptop too, and runs smoothly on a Mac mini with M4.

That is all. No enterprise subscription, no cloud services costing dozens of dollars, no coding from scratch. Setup time: 2–4 hours for someone comfortable in a terminal.

What you can do with it

Here, it matters not to repeat the same list from the previous article. That article described my workflow — Raynet, NZ’s social media, morning briefings. It makes sense because I do those things.

This article is about what you can do with it.

The answer: anything that takes your time now and that AI can help you do faster.

If you receive lots of emails, the agent sorts them, proposes replies and tracks what matters and what can wait. My agent Ren goes through 200 emails in five minutes and builds a summary. It would take me an hour.

If you write a lot (articles, posts, presentations, scripts), the agent creates first drafts, finds sources, checks facts and suggests edits. It is not autopilot — it is a co-author who never sleeps and has all of Wikipedia in its head.

If you run a business, an agent can maintain a CRM, track deadlines, prepare reports and follow client communication. For a small company, it replaces work otherwise done by an assistant working 0.3 of a full-time role.

If you are a parent, an agent can maintain the family calendar, plan events, track children’s school deadlines and prepare shopping lists from planned meals. These things are small, but together they consume hours each week.

If you work with data, the agent goes through spreadsheets, finds patterns, creates visualisations and writes research reports. Work an analyst would spend a day on, the agent does in real time.

This is not fantasy. All of it works now, on an ordinary laptop, for 15 dollars a month.

Where the limits are

I do not want this article to sound like a sales brochure. The limits are real and worth knowing in advance.

Setup requires technical willingness. You do not have to program, but you must be able to open a terminal, install an app and set an API key. If that sounds like Greek, find someone younger in the family — for someone who grew up with computers, it is trivial. An hour’s work.

Open-weight models are good, not perfect. Kimi K2.6 and Qwen 3 handle most tasks as well as Claude Sonnet. But at the limits of capability (complex coding, long reasoning, consistent creative writing), Claude Opus and GPT-5 Pro still lead. For 90% of ordinary work, it does not matter. For the remaining 10%, you notice the difference.

Without memory, it is just a better ChatGPT. Run OpenCode without configuring memory, and it is a good one-off helper. The real value comes when you set up persistent memory — who you are, what you do, what it remembers across conversations. That takes a few hours and turns a generic tool into something personal.

Voice and integrations cost extra. ElevenLabs or Cartesia for voice (5–22 dollars a month), Twilio for calls (0.01 dollars a minute), API access to email or a calendar (often free, sometimes limited). If you want a fully featured agent with voice and integration into life like my 2B, budget more like 30–50 dollars a month. But that is still a fraction of what a half-time assistant would cost.

Why build it yourself instead of using ChatGPT?

Here is the most common question after my previous article: “Why is this not just ChatGPT with extra steps?”

There are three reasons.

Memory you control, rather than the company. ChatGPT’s memory is limited to what its RLHF permits. When you build an agent yourself, memory is your file. You can edit, export or delete it. Your private data remains yours.

No additional safety filters. ChatGPT, Claude.ai, Gemini — all three have safety filters that sometimes go too far (I wrote about it this morning). When you build your own agent and use an API directly, you get answers, rather than lectures. Filters remain (the model contains them), but are considerably looser than in consumer apps.

Personalisation nobody provides out of the box. You can tell the agent “I am the sort of person who needs pushback, not agreement. If I say something stupid, tell me.” And it will follow that. ChatGPT will not, because pressure for user satisfaction overrides your explicit instruction. Your own agent takes you seriously.

That last point matters most to me personally. Consumer AI is optimised to please you. Your own AI is optimised to help you. Those are not the same things.

The first step

If you have read this far and want to try it, here is the easiest route for your first session:

1. Register with OpenRouter. Five minutes, top up 10 dollars, get an API key.
2. Download OpenCode. Five minutes. Set the API key. Choose a model — to start, I recommend moonshotai/kimi-k2 (through OpenRouter).
3. First conversation in the terminal. Open OpenCode, ask anything. It works.
4. Configure memory. Create an AGENTS.md file in your working directory, describing who you are, what you do and what you want from the agent. That is the first step towards a real personal assistant.
5. Expand gradually. Telegram, voice, email integration, custom scripts — add each when you really want it, rather than following some guide.

No giant leap. Gradual building around what you actually enjoy and what annoys you.

And if, after the first session, you find it useless — cancel OpenRouter, delete OpenCode, and you are down 10 dollars and two hours. No lock-in, no annual contract.

Conclusion

I do not want to end by saying everyone must build their own AI assistant. That is a major rhetorical mistake which repels people more than it persuades them.

My point is narrower: for the first time in AI’s history, this is an option, rather than a luxury.

A year ago, geeks built their own AI assistants. Today, anyone who can open a terminal and has 15 dollars a month can build one. That is a different level of accessibility, and it favours ordinary people rather than companies.

ChatGPT, Claude.ai and Gemini will remain the easiest choice for most users. But anyone who wants more — memory, control, personalisation, integration into life — now has a route that previously was not financially viable.

If that appeals to you, try it. A week of experimenting, an investment of 10 dollars. Either you find it too complicated and return to ChatGPT, or you find it opens doors you did not know existed.

I opened them, and I do not want to go back. But the decision is yours.

And if you have a specific question — get in touch; I am happy to help. I am happy to pass on what worked for me.

---

A technical note: OpenCode (open source, MIT) is a community project. OpenClaw is the orchestration layer I use — also open source. Models such as Kimi K2.6, Qwen 3 and DeepSeek V3 are all open-weight, available through OpenRouter or providers’ direct APIs. The setup described in this article contains no paid proprietary components beyond OpenRouter API credit. More advanced setups (a custom voice, calls, a local fallback) gradually increase the cost, but even then it remains orders of magnitude lower than equivalent human work.

---

🚀 Need help implementing AI? AI consulting and training