DEV Community

oVu X
oVu X

Posted on Originally published at github.com

Kwant Puts a Team of AI Bots on Your Own Machine

What it is

kwant is a free, open-source app that runs a team of AI bots on your own machine. It packs the whole team into one binary, so you host it yourself and keep your data at home. The project started on 22 October 2026, and it already has 689 stars on GitHub. Its author calls it an open-source, self-hosted option next to Grok Bot, Claude Duet, and OpenAI dots. This article checks that claim against the project README. kwant has no token, and its only official X account is @kwantbot.

How the team works

Each bot in kwant has a role, a voice, and a model. Bots hand work to each other with an @mention, and a team lead runs short talks and posts one summary. You can split your life into spaces, one for work, one for a side project, and one for home. Each space has its own bots, teams, channels, routines, accounts, and approval rule. A router called Laya picks the model and the effort for each bot set to auto. You can connect Claude, Codex, or Grok logins, API keys from Anthropic, OpenAI, or xAI, or your own server.

Done means checked

The core idea is simple: a bot must not mark its own work as done. Work shows up as cards on a Mission board, one card per hand-off. A card turns Verified only when kwant runs the check itself, such as a test or a file check. The card keeps the receipt, and files a bot attaches count as proof to look at, never as proof on their own. For code, the Dev desk shows what landed in each repo in plain words. A bot that did not write a pull request reviews it, tests and secrets included.

Facts, routines, and loose ends

kwant also checks facts with care. A bot backs each web fact with a link that opens the page at the exact quote. kwant checks that the quote sits on the page word for word, and it refuses a quote that is not there. Routines run on a schedule, and an Inbox lists what waits on you, what finished, and what failed. kwant can also read your chats, your mail inbox, and your GitHub issues to find work you started and never finished. It drafts replies for letters and bills, but it never sends anything on its own.

Stay in control

A team of bots that acts on its own needs a brake, and kwant builds one in. The default mode is Ask only when risky. Risky steps such as git push, network calls, unknown commands, or changes outside the bot folder show an approval card with the exact command. You press Allow once, Always allow, or Deny. Payments always ask, in every mode, even when approvals sit off. A mode called Flight School lets each bot earn more freedom per kind of work, but only you grant it.

How to start

The fastest start is the desktop app for Mac, Linux, or Windows. You download it, add an AI account you already have, and ask your first question. The Mac app is one dmg for Apple silicon and Intel, and Linux gets a deb or an AppImage. You can also build from source with Rust and Node.js 20 or newer, and a command called kwnt doctor checks the tools you need. A phone app installs from the browser with no app store, but it needs a kwant server your phone can reach over https. On a server, you can run the binary or use Docker Compose, Render, CasaOS, Umbrel, or Coolify.

Honest limits

kwant itself costs nothing, but the AI accounts behind it do not. You need at least one Claude, ChatGPT, or Grok account, or an API key, so model use still bills through your provider. On first start, the Laya router pulls its Python packages and its model from PyPI and Hugging Face, unless you turn it off with --no-laya. kwant sends no telemetry, and its screens load nothing from other servers.

The Linux and Windows apps are beta, and the Windows app does not update itself yet. Grok, the VM, and video checks are not on Windows yet, and Linux has no ARM app yet. The Mac app is not notarized by Apple, so macOS stops it the first time and you must click Open Anyway. The project is young: it started in late October 2026, and 689 stars is a small base next to the tools it cites.

Verdict

My verdict: kwant is worth a test run if you already pay for an AI account and you want your bots to run at home. The Verified check is the real idea here. It fixes the most common flaw in bot teams: a bot that calls its own work done. The approval cards and the payments rule give you a clear brake on risky steps. But treat it as a young beta, not as a finished product, and keep it off work you cannot redo. Start with one space, one small job, and watch the Mission board before you trust it with more.

Follow along

I test AI tools and build in public:

Top comments (0)