kwant

kwant is an open-source, self-hosted team of AI agents that you run on your own machine. The whole product ships as a single binary, called kwnt, which you can install on a laptop or a VPS and which brings the app and the database with it. Instead of a single chat tab, you get a group of bots, each with its own role, voice and model, that share chats, pass work to each other with @mentions and keep going on a schedule even after you close the tab. It is built for people who want an AI team that proves its work rather than simply announcing that a job is done. Most AI tools answer a question and then declare the job finished, leaving you to take the model's word for it. kwant takes a different position: chat apps and agent frameworks can do the work, but they rarely prove it. In a typical chat assistant there is one assistant per chat, one answer is picked for you, and when facts conflict you only ever see the version that was chosen. Agent frameworks can be wired up to do better, but you have to build the routing, the evaluations and the approvals yourself. kwant is built around the idea that a result should be checked, that a fact should open at the exact quote it came from, and that nothing should be paid without you. Each bot has its own role, voice and model, and knows who else is in the room. You hire bots from 27 roles, and each message goes to the bot that fits, with kwant picking the bot, the model and the effort and showing you why. When one bot finishes its part it hands the work on with an @mention: in the product's own example, TechLead takes a request to ship a pricing page and passes the layout to Designer, who then asks QA to check the layouts on a phone. Team discussions work the same way, with the lead running a short discussion and posting a single summary, while the full fold of every turn stays available to read. New teammates are added from presets such as Security, Research or Chef, and the name, role, avatar and voice fill in automatically. Verified results are the core of the product. Bots split work into cards, and a card turns Verified only when kwant runs the check itself, so "done" means it was actually checked. Web facts come with links that open at the exact quote, and no quote on the page means no claim. When two facts conflict, they wait for you: the Truth Ledger shows both and you keep one. Goals let you state what should stay true, kwant checks it, and when a goal breaks a bot offers one fix for you to approve. Smart Results turns a table or list in a reply into a comparison you can sort, a chart or a checklist, copied word for word. Risky steps are handled deliberately. Ordinary work just runs, but something risky, like a push, a network call or a change outside the bot's folder, shows you exactly what will happen first, with the full arguments in front of you and Allow once, Always allow or Deny as the options. That prompt can be switched off too. Payments always ask, even when they sit behind a plain "Continue". Autonomy is earned rather than assumed: in Flight School bots earn tiers per kind of work, only you grant them, and new bots start at Act. AutoFix, audits, Flight School and the Truth Ledger all run without setup. kwant is designed to keep going when you are not watching. Say "every morning" and the bot sets up the routine itself, with Undo, Edit and Pause on the card, or you can create one for daily, weekly, every few minutes or a cron rule. The Inbox page collects what happened while you were away on one screen: what is waiting on you, what finished, and what got stuck. Loose ends finds the reply you owe across chats, email and GitHub and drafts it, though it never sends. Letters lets you photograph a bill or notice so kwant can pull out the deadline and amounts and remind you. Accounts and spaces keep everything separated. You can use Claude, Codex or Grok logins you already pay for, API keys or your own server, and when one account runs out of usage the turn moves to the next in your order across Claude, Codex, Grok and Custom API, with the chat saying so. Spaces give work, side projects and home their own bots, chats, routines and accounts so nothing leaks between them; each space gets a name, an icon and a colour, switches with Command or Ctrl plus 1 to 9, and opens where you left it. Getting started requires Rust and Node.js to build it plus one AI account. You run make run, the web app opens on your machine, and on a server you open the link it prints once. Accounts are added by using a CLI you are already signed in to or by pasting a key, and you can add as many as you like. There are downloads for Mac, as a universal build for Apple silicon and Intel that is not yet notarized by Apple, and for Linux as a .deb for Ubuntu and Debian on x86_64 or an AppImage, plus a Docker option. kwant is AGPL-3.0 open source, at version 0.6.8 with 19 releases since October 25. Typical scenarios follow the shape of a working day. A research bot runs every weekday at 07:00, sums up three new app reviews, spots a price change on a rival's site and puts a draft reply in Loose ends for you to review the next morning. A team ships a pricing page by handing the layout to a designer and the phone check to QA. Parallel Universe load-tests up to three ways to deploy on your own server before you pick one. You approve an action from your phone away from your desk, or photograph a notice to have the November 5 deadline pulled out and tracked. kwant is aimed at people who want an AI team on their own hardware without writing code. It works with Claude Code, Codex, Grok, API keys, any MCP server, Ollama and any OpenAI-compatible server, and runs on Mac, Linux and Docker. It is free and open source, and you bring the model accounts you already have. The takeaway is straightforward: kwant gives you your own self-hosted team of AI bots that share chats, hand off work, check each other, quote their sources and ask before they spend, all from one binary on your laptop or server.