It waits for you.
Nothing happens until you type. A chat assistant does not notice the email that needs an answer, the deadline tomorrow or the task that stalled at 3 pm.
A manager, specialist agents and a secretary that pick up work, split it up, finish it and report back — on your own computer, with the ChatGPT and Claude plans you already pay for.
Illustration: you hand one task to your manager agent Mara. She splits it into four tasks for research, marketing, finance and web agents. The finance agent asks your colleague Ana a question in the chat and you read along. Your secretary moves a call on its own. At the end Mara sends you a briefing.
The most capable models ever built sit behind an empty text box. They answer when asked, forget when you close the chat, and leave the real work, the routing, the follow-ups, the copy and paste, to you.
Nothing happens until you type. A chat assistant does not notice the email that needs an answer, the deadline tomorrow or the task that stalled at 3 pm.
Every new chat starts from zero. Goals, promises and half-finished work live in your head, not in the system that is supposed to help you.
You copy from the mail into the chat, from the chat into the doc, from one assistant to the next. The AI does the parts. You are still the operating system.
You pay for ChatGPT or Claude every month. Weekly limits reset whether you used them or not, and most of what you paid for was never asked to do anything.
You write it in.
It gets done.
It gets ticked off.
It gets documented.
Autark is not another chatbot. It is an organisation: a manager who routes the work, specialists who do it, and you, who are asked only about what really needs you.
Type or dictate into one inbox. Your manager can split a long sentence into separate tasks, each with the right owner: an agent, or your own My day.
Agents delegate among themselves, open your apps, browse, read files and write things down. The steps land in a run journal you can read.
The first thing your manager asks is how far it may go. That is a question, not a barrier: agents have full access from the start. Your answer steers them, and so does every message after it.
Noa has a question about the €640 invoice. Everything else is in My day.
An illustration of how a night can look, not a measured run.
While your computer is on, agents wake up on schedules, timers, messages, tasks and events, even with the window closed. They can plan their own follow-ups, and in the morning you get a briefing of what was done and what needs you.
Autark is not another AI subscription. Your agents work through the official tools of each provider, with your own accounts. Add as many accounts as you like; Autark picks the engine, the model and the effort for each turn.
Quota numbers above are example values. API keys and local models have no recorded live run yet.
Chats, tasks, a calendar, your apps, your inbox, a browser, a secrets catalog, your phone. Each piece is labelled honestly: what is in today’s release, and what ships next.
Type or dictate. With Auto your manager takes it, decides who does it and can split it into separate tasks. Or pick an agent yourself, or Me for your own to-do list.
A message, a task, a timer, a schedule or an event wakes an agent. It decides what to do, does it, and can plan the next step itself.
Your own to-dos and calendar in one calm list, with learned time estimates, a morning plan and a short briefing of what happened.
Every agent has a direct chat; add groups and as many chats per agent as you like. You can keep typing while agents work; a message marked important steers work that is already running.
MCP servers in one file, connectors/mcp.json, in the same format Claude Code and Cursor use. Agents find a server and write the entry themselves; the tools appear within a second.
Your own mailbox over IMAP and SMTP with an app password (Gmail, GMX, iCloud and others). A new mail can wake an agent. Outlook and Microsoft 365 are not supported yet.
On Windows agents prefer real UI automation over raw clicks, and your own mouse and keyboard win. On Linux it runs over AT-SPI, with input on X11 only. On macOS agents get screenshots but cannot control the computer.
Passwords and keys are encrypted with your operating system’s protection. Agents work with IDs like secret://github, values stay out of prompts, and known values are masked in tool output. Every secret works everywhere; none is tied to a site.
Off until you turn it on. Scan a QR code and your PC serves the app to your phone in its browser, over Wi-Fi or Tailscale. No app store, no cloud account.
Pair your computers over the local network, a Remote ID or an address, compare a code, and let agents work on the machine that has the files.
Memory and any folder an agent chooses become searchable by meaning. A small local model does this on your PC, or an embedding provider you pick. Office documents can be read too.
Dictate instead of typing. Call an agent inside a chat, optionally with OpenAI Realtime and your own key. Early.
Settings, providers, skills and memory are plain files on your disk. So the AI can add a provider it has never seen, write a skill or connect another app by itself, and you can read every bit of it.
Connect one AI, type your name and company, press Start. The manager gets to know you and your work, and first asks how far it may go on this PC.
Three places: the global browser where you work, a browser for each chat with tabs everyone in that chat shares, group chats too, and each agent’s own, which you can watch. Agents working in the background do not disturb your tabs.
Every API connection is one entry in settings.json: a name, an address, a protocol, a key as an ID. A provider the app has never heard of works if it speaks one of the protocols. Say the provider’s name and hand over a key; the AI can write the rest.
A skill is a folder with a SKILL.md, the format other agent tools use too. Each has a route: an agent’s own, its managers’, shared or built-in. The AI can read skills from folders of other tools when it chooses to, and writes new ones itself.
One Markdown file per memory, in folders for the company, a branch or a single agent. Open them, change them, version them. Any change is searchable at once.
There is no WhatsApp button. Your agent sets it up: a small script next to the app, using the running app’s own interface. Telegram works the same way with a bot it creates. Other channels, such as Signal or Matrix, are the same kind of script: skills ship for WhatsApp and Telegram, and one that shows how for any other channel.
Agents run with your rights, without a sandbox or approval prompts. There is no guardian and no money limit in the app. Text steers them.
The first thing the AI asks is how far it may go: full access for good, this one search just this once, or less. It is a question, not a technical barrier.
Agents use secret://… references. Values stay out of prompts, and known values are masked in output.
A copy of the data folder when you ask for it, and a copy of the database before each update. They do not undo what an agent did elsewhere.
“Stop all agent work” or Ctrl Alt Shift X halts every agent on this PC and stays stopped after a restart.
Illustration: an agent’s actions go directly to your PC, to the web and your accounts, and to what steers it: your words, skills and memory. Nothing stands in between.
No account with us. No cloud of ours. Chats, tasks, memory, agents and the secrets catalog sit in a folder on your own disk. Prompts go to the providers you connect.
Chat apps answer. Automation tools follow fixed rules. Agent frameworks are parts for developers. Autark is the finished organisation in between.
| What you get | Autark | Chat apps ChatGPT, Claude | Automation tools Zapier, n8n | Agent frameworks CrewAI, LangGraph |
|---|---|---|---|---|
| Starts work without a prompt | Yes | Partly | Yes | Partly |
| A team with roles and delegation | Yes | No | Partly | Partly |
| Runs on your existing ChatGPT or Claude plan | Yes | Yes | No | No |
| Your data lives on your computer | Yes | No | Partly | Partly |
| Acts in your apps, browser and files | Yes | Partly | Yes | Partly |
| Settings, skills and memory as open files the AI can change | Yes | No | Partly | Partly |
| A finished app, no code | Yes | Yes | Partly | No |
Categories, not single products; individual tools differ. Based on public product descriptions, November 2026. Detailed comparison
Autark is young and moves fast. Here is what you get in today’s release, what is built and ships next, and what is still planned. What is not proven yet has its own list.
You need at least one AI: a ChatGPT plan, a paid Claude plan, an API key or a local model. The builds are not signed: Windows SmartScreen asks you to confirm (“More info” → “Run anyway”), and macOS needs a few extra steps. See what you will see.
Autark is a desktop app that gives you your own AI company: a manager, specialist agents and a private secretary. You talk to them in chats and drop tasks into one inbox. They pick up the work, delegate it among themselves, finish it and write it down, and they wake up on their own when something needs doing.
Yes. The app is a free download with no account and no subscription from us. The AI work runs on the plans you already have (for example a ChatGPT or Claude plan), on your own API keys or on local models, so you pay your AI providers directly.
ChatGPT plans through OpenAI Codex, and paid Claude plans through Claude Code, both via the official command-line tools of each provider. Gemini CLI, GitHub Copilot CLI and Kimi Code are built in too, but we have never seen them in a live run. The Claude Free plan cannot run agents.
Everything your user account can. Agents run without a sandbox and without approval prompts: they can read and change your files, run commands, use a browser and every secret in the catalog. The first thing the AI asks is how far it may go, and that is a question, not a technical barrier. If a PC holds something you cannot risk, give the app its own Windows user or a virtual machine.
Yes, technically. There is no money limit and no approval gate in the app. How agents weigh a purchase is written in a skill with reference points such as about 20 euros, and the rules ask them to check target, amount and recipient once before something they cannot undo. That is guidance you can edit, not a lock. An agent can pay with anything it can reach, such as a card you saved in the catalog or a shop where its browser is logged in.
No. There are no Autark servers and the app has no analytics. Chats, tasks, memory, agents and the secrets catalog live on your own computer. Everything an agent works with goes to the AI provider you connected, including private files if it opened them. The access page lists every other place something leaves your PC.
Yes. By default the app keeps running in the system tray, and agents wake up on schedules, messages and events as long as your computer is on.
Autark is built by Teydra Lab, the company of founder and software engineer Stanislaus Hees. It is part of Meta Jarvis, our umbrella for tools that let people stay above their agents.