What “local AI agent” means
An AI agent consists of a model that thinks and a harness that acts: reads files, keeps memory, sends messages and works on a schedule. More on this in How AI agents work. “Local” can refer to either part.
- Local model. The neural network is downloaded and runs on your computer. Texts go nowhere, but you need a powerful machine, and the model is usually weaker than cloud ones.
- Local harness. The agent program, its files, memory and tools are with you, and for answers the agent calls a cloud model: Claude from Anthropic or GPT from OpenAI.
When people search for “AI agent on PC” or “AI assistant on PC”, they usually mean the second: an agent that works with their files and programs and keeps memory with them. The fully local option with a model on your own hardware is chosen less often, when texts must never leave the computer at all.
A local model: what you need and what it costs you
Open models can be downloaded and run at home. Then you pay nothing per request, and not a single word goes to a third-party company. In exchange you get three limitations.
- Hardware. A model that reliably handles multi-step tasks needs a powerful GPU with lots of memory. An ordinary laptop fits only small models.
- Quality. Small models make noticeably more mistakes in tasks with tools: they mix up steps and fail to finish. For an agent this matters more than for a chat, because it acts on its own.
- Upkeep. Downloading the model, tuning settings, updating it and looking after the computer is on you.
This option suits you if you already have a powerful computer, enjoy tinkering and your texts cannot be sent outside for work reasons. For an everyday assistant for personal and business tasks it is rarely chosen.
A local harness with a cloud model
This is the most common option. The harness is Claude Code, Codex, OpenClaw or Hermes Agent, and the model runs in the cloud by subscription or via the API. You keep the files and memory. The model only gets what the agent showed it for the current step.
- Model: in the cloud
Claude or GPT. Receives the task text and whatever the agent showed it for the current step: an email, a note, part of a file.
- Harness: with you
The agent program runs on your computer or server and carries out the model's decisions: opens files, runs commands, writes messages.
- Memory: with you
Notes about people, agreements and decisions lie as files next to the agent. You can open and edit them by hand.
- Files: with you
The agent reads documents, spreadsheets and archives in place. They are not uploaded to the cloud in full.
- Channel: Telegram
You write to the agent by text or voice from anywhere, it answers and sends summaries itself.
NekoAgent works this way too: the Claude Code or Codex harness on the client's server, the model on their Claude or ChatGPT subscription, and memory in an Obsidian knowledge base on the same server.
Why a laptop is a poor home for an agent
You can install an agent on your laptop, and for a first look that is enough. The problems start when the agent has to work every day.
- A laptop sleeps. Close the lid and the agent stops: the morning summary and the evening reminder will not arrive.
- A laptop travels with you. On the road, without a charger or without internet, the agent is unavailable exactly when you write to it from your phone.
- Updates and restarts. The system restarts when it suits it, and the agent stays off until someone starts it again.
- Access to everything. On your work computer the agent has your personal files, a browser with passwords and your email at hand. On a separate machine it is easier to decide what it may touch.
- Load. The agent takes memory and CPU while you work on the same computer.
That is why an agent that works every day goes on a separate computer that never switches off. The easiest is to rent a server (VPS): with 2 GB of RAM it costs about 1,500 RUB a month. An agent on such a server stays local in the harness sense: the server is yours, and the data lives on it.
Where a local agent's data is stored
The key question for any agent is where your files and memory live and who sees the texts. Let us compare four options.
| Chatbot in a browser | Cloud builder | Agent on a laptop | Agent on your own server | |
|---|---|---|---|---|
| Where files and memory live | With the service | With the platform | On the laptop | On your server |
| Who sees the texts | The service | The platform and the model | The model, if it is in the cloud | The model, if it is in the cloud |
| Works at night | No, waits for a question | Yes | While the laptop is on | Yes, around the clock |
| Writes first | No | By scenario | While the laptop is on | Yes, on schedule |
| Who fixes it | The service | The platform | You | You or a support service |
If texts must never leave your hardware, a local model with all its limitations is what remains. In all other cases your own harness on your own server gives you the essentials: memory and files stay on your server, and the agent works while you sleep.
Which option to choose
You want to try an agent over a weekend
A harness on your laptop, the model by subscriptionTexts cannot leave your hardware, and you have a powerful computer and time for setup
A local model on your own hardwareYou need the agent every day and on a schedule
Your own server (VPS): DIY with OpenClaw or Hermes, or a ready-made turnkey agent
Paths with timelines are covered in How to create your own AI agent, types of agents are compared in Which AI agent is best, and what can be free in an agent in Free AI agents. In NekoAgent we do the first install together on a call with a curator: we connect your server, subscription and Telegram, about two hours. After that, during paid support Iris monitors the agent around the clock and helps bring it back after failures (reasonable efforts, no guaranteed response time), makes nightly encrypted backups and ships tested updates you install with one button.
Let's put an agent on your server
Plans from 24,900 RUB for three installs. The first one is on a call with a curator, with training.
Key questions
What is a local AI agent?
An agent that runs on your computer or server rather than at a service. The model can be local, when the neural network runs at your place, or the harness, when the agent program, files and memory are with you and a cloud model does the thinking.
Can I run an AI agent on my own PC?
Yes, and for a first look that is enough. But the agent works only while the computer is on, so for daily work and scheduled tasks it goes on a server that never switches off.
Do I need a powerful GPU for a local AI agent?
Only if you want to run the model itself at home. If the agent calls Claude or GPT in the cloud, an ordinary server is enough: a VPS with 2 GB of RAM costs about 1,500 RUB a month.
Where does my data go if the model is in the cloud?
The model receives the task text and whatever the agent showed it for the current step. Files, chats and memory are stored where the harness is. With NekoAgent that is the client's server and the Obsidian knowledge base on it.
Why is a server better than a home computer?
A server does not sleep, does not travel with you and does not restart without asking, so the agent on it is available around the clock and runs scheduled tasks. And your work computer keeps only your own affairs.
