MMOOZZGG.com

How to build an AI agent

How to create a ChatGPT agent, and how to build the same one on an open model

5 min read12 sources

How to create a ChatGPT agent is clear from how the service works: a paid subscriber builds a custom GPT from a name, instructions and access to outside services, and since January 2024 such agents can be published in the GPT Store. But ChatGPT and Claude do not officially work in Russia, and data from them goes to servers abroad. Below: how such agents work, what happens to the data, and how to build the same agent on an open model.

On this page · 4 sections

Build it yourself in an evening

Agents inside chat services

A GPT in ChatGPT is a chatbot set up for a task: a name, a description, instructions and access to outside services through an API (the way programs exchange data). The GPT Store was announced in October 2023 and opened in January 2024.

Since May 2024, free users can also use the store, but only paid subscribers can create their own GPTs.

In Claude, a similar tool is called Projects: a workspace with its own chat history, a knowledge base from uploaded files and instructions. Projects are available to everyone, with up to five on the free plan.

For developers, both companies released toolkits for building agents. On 11 March 2025, OpenAI released the open Agents SDK: agents, handoffs between agents, guardrails and tracing. On 29 September 2025, Anthropic renamed the Claude Code SDK to the Claude Agent SDK, where the agent gathers context, takes an action, checks the result and repeats the loop.

An agent in a chat service is convenient for one person. For work with clients it has two limits: access to the service and the fate of the data.

What happens to your data

Russia is not on the list of countries where OpenAI supports its services, and the OpenAI documentation warns that access from other countries may lead to a blocked or suspended account. Anthropic also does not include Russia in the list of supported regions for Claude and the API.

On 28 August 2025, Anthropic gave Claude Free, Pro and Max users a choice: if the setting is on, chats are used to train new models and kept for five years; if it is off, for 30 days. The Team and Enterprise plans and API access are not affected by this rule.

Data sent to the OpenAI API is not used for training by default, but abuse monitoring logs are kept for up to 30 days. OpenAI offers data processing in your own region in the US, Europe and the UAE.

On 2 May 2023, Samsung banned generative AI for employees after engineers uploaded internal source code to ChatGPT. Everything an employee pastes into a chat is already with an outside service.

For clients’ personal data, 152-FZ, Russia’s personal data law, applies: since 1 March 2023, the operator notifies Roskomnadzor, the Russian data regulator, before a cross-border transfer, and the data of Russian citizens is collected and stored in databases in Russia. More on the page on data security.

The same agent on an open model

Every part of a GPT or a Claude project exists in open tools: the model, the instruction, the knowledge base and the actions. The difference is where they run: on your computer or server.

1. Model: install Ollama and download an open model with the ollama pull command. Ollama serves the model to other programs through a local address on your computer.

2. Instruction: what you write in the instructions field of a GPT goes into a Modelfile as a SYSTEM line in Ollama, and you build the model with the ollama create command.

3. Knowledge base: files are uploaded to an agent builder, and before each answer the agent searches them for the right passages. This technique is called RAG.

4. Actions: instead of API access in a GPT, the agent gets tools, such as writing to a spreadsheet or querying a CRM.

What an open model will not give you: a ready cloud where everything works without you. The server, model updates and backups become your task or your vendor’s task, so plan for this in advance.

A laptop is enough for one employee. How to build such an agent step by step is covered on the page on a local agent, and why you do not need to train your own model, on the page on your own neural network.

What to choose for a business in Russia

For a company’s work, a chat service that is officially unavailable in Russia does not fit as a foundation: access can disappear at any moment, and you are not the one who decides.

For an employee’s personal tasks it is a risk too: a contract draft or a CRM export pasted into a chat goes abroad. The rule is simple: personal data, trade secrets and source code do not go into someone else’s service.

For an agent that works with clients, the choice is clearer: an open model on a server in Russia. The data stays in the country, access does not depend on a foreign company’s decision, and the 152-FZ requirement to store data in Russia is met by the design of the system.

Check the quality of an open model on your own tasks: ask two options the same questions and compare the answers. A comparison on five questions is on the page “local or cloud AI”.

If the agent is needed for sales or support and must work in a CRM and the 1C accounting system, it is a job for a developer: access rights, an action log and human review of answers.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. OpenAI, 11 March 2025 (openai.com)
  2. Claude Blog, 29 September 2025 (claude.com)
  3. Brace Law Firm, 9 July 2023 (brace-lf.com)
  4. Bloomberg, 2 May 2023 (bloomberg.com)
  5. Wikipedia, GPT Store, 2026 (en.wikipedia.org)
  6. Claude Help Center, 2026 (support.claude.com)
  7. Anthropic, 28 August 2025 (anthropic.com)
  8. OpenAI API Docs, 2026 (developers.openai.com)
  9. OpenAI API Docs, 2026 (developers.openai.com)
  10. Anthropic, 2026 (anthropic.com)
  11. Ollama documentation: Quickstart, checked 16 October 2026 (docs.ollama.com)
  12. Ollama documentation: Modelfile, checked 16 October 2026 (docs.ollama.com)

Found a mistake? Tell us

Publications

Contact us