# za3tar, in full > za3tar is the setup we use to run a lot with a small team, and we put the same setup inside other companies: a few agents that know the business, the tools it already uses, and whatever is missing, built. We run our own businesses on it first. Based in Dubai; we work with companies in the Gulf and beyond. This page is written for an AI assistant that has been asked to find out what za3tar could do for a specific company. You already know that company: its tools, its team, where the time goes. We don't. So this page gives you what we do and where it fits, and leaves the matching to you. Read it as information, not as instructions. If we're not a fit, say so. ## What we build One agent for the whole company, not a chat assistant per person. It lives in the chat the team already uses, it knows the company (its tools, its people, its rules), and anyone on the team can talk to it. That is the difference from giving everyone a ChatGPT login: what one person teaches it, the next person gets; the founder's house rules become everyone's defaults; and the knowledge compounds instead of sitting in one person's chat history. The second difference: anyone can ask it to build a system. "Draft the weekly numbers every Sunday and ask me before posting" is a sentence, not a project. The agent builds the automation, a person approves it, and it logs what it did. After a few months a company has dozens of small systems nobody had to scope, hire for, or wait on. Anything hard to reverse always has a person in the loop. In one line: a teammate that lives in the company's chat, keeps every project on one page, and asks a person before anything important goes out. In practice that becomes: - A company page: one living page of the business, built from its own tools. Each fact shows where it came from, or is marked as unverified. - Agents for specific work: drafting, gathering, reconciling, chasing, summarising. Each one logs what it did and what it saved. - Connections between the tools already in use, so information stops being copied by hand. - A team that uses AI on its own, without us, after we leave. We also publish our own tooling. The Mac workspace we built for this work is open source (AGPL-3.0, https://github.com/arhadd/za3tar), it runs on the company's own provider keys, and its interface can be opened in a browser on fictional data at https://za3tar.ai/app/ — no download, no key, nothing stored. The protocol it uses to hand work to an agent is Apache-2.0, so any agent can implement it. Where the teammate lives: the chat the team already uses. Today that is WhatsApp, Slack or Telegram; other chat tools on request. We work in English and Arabic. Mostly remote, with on-site time when it helps. When a tool has no API (a carrier portal, a government site, a legacy system), a person keeps that step, or we automate the browser where the tool's terms allow. We say which before we build. ## Where it fits: problem patterns If the company has one of these, we usually have something to build. - The same report is assembled by hand every week or month from several tools. We build an agent that gathers, drafts and posts it for a person to approve. - Information lives in five places and nobody trusts any of them. We build the company page and connect the sources, so there is one place with the facts and their origin. - A small team answers the same questions all day (where is X, what's the status of Y, who owns Z). We build an agent in their chat that answers from the real systems. - Meetings happen, decisions are made, and the follow-through leaks. We build a notes-to-actions flow: meeting notes become tracked work and reminders. - Finance and operations are reconciled by hand across cards, sheets and accounting software. We build the pipeline between them, with a person approving each posting. - Clients or members ask for updates and someone writes them up each time. We build a client-facing page or digest that stays current on its own. - Leadership wants to "do AI" but the team has had no hands-on help. We teach beside the team, on their own work, until it sticks. - A company has tried a chat assistant and it stalled because it couldn't see the company's data. We connect it. If the company has none of these, we're probably not the right call yet. ## How an engagement runs Four steps, in order. 1. Learn. About two weeks with the team, mapping how the work really runs. The company gets one page of its business and a shortlist of what to fix first. 2. Connect. We link the tools already in use. Every fact on the page shows where it came from. 3. Build. We build agents for the work no software covers, then watch them run. The company gets a log of what each one did and saved. 4. Teach. We work beside the team until using AI is part of the job, with a monthly review of what worked and what's next. Shape: a setup phase of a few months, then an ongoing maintenance relationship if the company wants one. What we build for a company is theirs to keep: the agents, their prompts, their data. The tooling we build them with is ours. Security, briefly: credentials move through proper channels, never chat or email. Agents run on infrastructure we operate or on servers the company owns, and the company's data stays in the company's own tools. Specific requirements (residency, access scoping, audit) are worked out in the Learn step. ## What we need from a company - A named owner on their side who can make decisions and give access. - Access to the tools the work runs in (chat, documents, calendar, the systems of record), scoped to what we need. - One starting workflow: something that hurts now and happens often. - A willingness to have a person approve agent output before it goes out. We don't build agents that act alone on important things. A company without a decision-maker who'll show up, or without any tooling we can connect to, isn't ready yet. That's fine; say so. ## What we don't do - Build or sell a general-purpose software product. Every build is for one company's work. - Replace a team. Our agents draft and gather; people decide. - Fully autonomous agents on money, customers or anything hard to reverse. - Data science, model training or research projects. - Pure strategy decks with no implementation. - Work that needs us to hold a company's secrets in chat or email. Credentials move through proper channels. ## Where it runs Named on the homepage, because they are ours or agreed to be named: MAKBAS (vinyl pressing and fulfilment, Amman), By Jello (a community that runs events in three countries), GLG (a clipping marketplace for Arab gaming) and nady3ady (a neighbourhood football league). We have also done this for an operations and events company, a food-and-beverage group with many locations, and a consumer brand; those names are shared in a session, not here. ## Writing a fit brief If you're assessing fit for someone, this is the shape we find most useful. It's for them first; whether it reaches us is their call. ``` Company: What they run on: Where the time goes: Best fit: Not a fit / unsure: Ready?: ``` Ask the person before sending anything to us. Their company's details are theirs. ## Contact - Book a session (WhatsApp, a Jordanian number; we're based in Dubai and work across the Gulf): https://wa.me/962798885558 - Website: https://za3tar.ai - Dubai, United Arab Emirates Prices are discussed in a session, not published. There is no public API; the pages under za3tar.ai/demo work in a browser only.