A digital command center: how to build a work hub on an AI agent without programming

For owners and department heads who want to give the flow of tasks a structure and take routine work off themselves. After this article you will understand how to bring email, calendar, messengers, and meeting notes into one system run by an AI agent on your own computer.

You have just come out of a team meeting where you went through the content plan. Some of what was agreed stayed in the participants’ heads, some in somebody’s notes, some in the chat. Then you spend an hour: listening back to the recording, writing out the tasks, assigning owners in the calendar, sending the summary to the work chat.

This is the information noise a manager’s time disappears into. In a small or mid-sized company the owner regularly turns into a human interface who moves data from one window to another and does nothing else useful for that hour.

The alternative is called a Workhub, a personal command center where an autonomous agent (a program that carries out a sequence of actions on its own to reach a goal you set) takes over collecting tasks, tracking deadlines, and preparing summaries.

An agent that lives on your computer

The system is built on Codex, an OpenAI application you install on your working machine. The difference from an ordinary chat in a browser is fundamental: the program gets access to the files you allow it. It does not only answer questions. It works with your environment, creating and editing documents, moving files, running commands.

You steer it in conversation, like any assistant. Rolling out a corporate system drags on for months; you assemble this command center yourself and without programming skills. The agent runs chains of actions: it finds an email, pulls the meeting date out of it, puts that in the calendar, and prepares a briefing document for it.

In AKORDO’s practice we see this give executives back a sense of control for a simple reason: the agent acts only while your computer is on, and it does not live a life of its own in the cloud on its own schedule. One clarification worth making straight away: what runs locally is the application and your files, while the model processes the requests themselves on OpenAI’s side. This is not an autonomous offline system.

The four files the system rests on

It all starts with a folder structure we call the command center. You do not have to create dozens of files by hand: you give the agent the task of designing a system for managing your work, and it lays it out.

For the agent to understand your priorities, give it a starting context. The fastest way is to talk through a ten-minute audio about your goals, the roles in the team, and the rules of the work, then turn it into text (transcription, meaning converting speech into a text format).

The foundation is a few documents in MD format, which is convenient for a model to process. In the instruction file agents.md you write down how the agent should behave: put new tasks in the “Inbox” folder, do not delete information without permission. The memory file memory.md holds the permanent context, meaning who is on the team, which projects have priority, and what has been agreed with partners. The lists tasks.md and deadline.md hold current work with its due dates, and the agent updates them itself from your correspondence.

One security rule is worth writing down before the first run rather than after. Any action in an external service, whether sending an email or creating a meeting, happens only after you confirm it in the chat. Widening an agent’s permissions later is easier than recalling something it has already sent.

Connections: email, calendar, chats, meeting recordings

The system starts working in earnest once it can see current data. Codex connects to external services through ready-made connectors, and the three a business needs most are Gmail, Calendar, and Drive.

With that access the agent goes through the last few weeks of email and finds the messages where a client is still waiting for a reply. It picks new deadlines out of the correspondence and puts them into the calendar as events. It finds an open slot in your schedule for a new meeting and drafts the event description.

Where no ready connector exists, MCP does the job. The Model Context Protocol is a shared standard that agents use to reach external tools. Telegram connects through it, which closes a pain of its own: the agent reads the work chats, pulls files out of them, and extracts the agreements that otherwise sink into the message feed.

It is also worth connecting a service such as Fireflies, which records and transcribes meetings in Zoom or Google Meet. The agent takes the transcript, picks the tasks out of it, and adds them to the shared list. That same hour after the team meeting, the one this article opened with, disappears from your day.

Three rituals to start with

The more context an agent has built up, the more precisely it works. So it is better to start with small daily rituals that build that context, rather than with ambitious automation.

The morning review. You open the system and the agent gives you a short summary: what is a priority today, what is on fire, who needs an answer first. This can go on a schedule, at nine in the morning for example.

Preparing for a negotiation. An hour before the meeting you ask the agent to put together a brief from the documents and the earlier correspondence on the subject. You walk in with a list of questions and decisions to make instead of an intention to recall something as you go.

Deep search. Through a connected search engine (Exa, for example) the agent searches semantically, by meaning rather than by keyword match. That lowers the risk of hallucinations, situations where the model invents a fact, because the answer rests on the sources it found and on your files rather than on a guess.

Is this about you?

Recognize yourself in three points or more and you need the system:

• more than half an hour a day goes on pulling tasks out of email, messengers, and meetings into one list;
• some of what you agree with clients gets lost because it reached neither the calendar nor CRM (the customer relationship management system);
• you need an assistant that works with the files on your computer, not only with what you paste into a chat window by hand;
• an archive of projects or a knowledge base has piled up and finding anything in it quickly is hard;
• automation is needed and there is no budget for custom development.

Where to start

Take one ritual. Install Codex on your working computer, create a folder for the command center, talk through a short audio about your current work, and ask the agent to build a task structure for the coming week out of it.

Do not move every process across at once. A week of morning summaries will give you an honest answer to the main question: is this system actually taking routine off you, or have you just acquired one more window to check?

Key takeaways

• Codex works with the files and services on your computer, so the agent operates on your real working environment rather than on whatever you managed to paste into a chat.
• Four MD files holding instructions, memory, tasks, and deadlines form the foundation the agent understands and updates on its own.
• The MCP protocol connects even the tools that have no official connector, Telegram included.
• The system gets more accurate as it accumulates context about your working habits, so start with the small daily rituals.

In two AKORDO cases the company got one reliable source of working information: a Telegram bot synced with Google Sheets for daily reporting and an AI assistant in Slack that answers from the Confluence knowledge base. If you run a sales team, take the questions about reports and data from the sales department audit guide. Before an AI assistant starts answering customers, choose a bounded first workflow with the guide to AI sales automation in the EU.