Ask once, the team passes it on
Bots share chats and hand work to each other with an @mention. You get one answer back.
Open-source, self-hosted AI agents in one binary
Your own self-hosted team of AI agents
They share chats, pass work to each other and get it done on their own, on a schedule or while you're away. It runs on your laptop or a VPS.
Free, for Apple silicon and Intel. Also for Windows and Linux, in beta.
Free, for 64-bit Windows. It isn't signed yet, so Windows warns the first time: click More info, then Run anyway. Also for Mac and Linux.
Free. A .deb for Ubuntu and Debian on x86_64; other distros, the AppImage (how to install). Also for Mac and Windows.
First time on a Mac?macOS asks once, because kwant isn't notarized by Apple yet.
- Drag kwant to Applications and open it. When macOS says it couldn't verify it, click Done.
- Open System Settings › Privacy & Security and scroll down to Security.
- Click Open Anyway next to “kwant” was blocked, then Open Anyway again, and enter your Mac password or use Touch ID.
You only do this once. On macOS 14 or earlier, right-click kwant in Applications and choose Open instead.
SoonA one-line Terminal install that skips this step.
0.9.4 released Nov 12 40 releases since Oct 25 AGPL-3.0, open source
Added in 0.9.0
Review a bot's work in one placeA second checker runs the checks again, and the Inbox shows the verdict, what it couldn't check, and Approve, Send back or Roll back.
Point at the part of the page you meanIn Watch live, click part of the page and the bot gets its HTML, size, styles and a picture.
Notes on diff linesLeave notes on lines of a Dev desk review and send them to the bot as one message.
A worktree for each bot taskMake a worktree from a branch for a bot or for you. Remove never deletes the branch.
Also new: roll back, even after you approved it, and watch a mission. What's new in 0.9.4 · in 0.9.3 · in 0.9.2 · in 0.9.1 · in 0.9.0 · in 0.8.2 · in 0.8.1 · in 0.8.0 · in 0.7.7 · in 0.7.6 · in 0.7.5 · in 0.7.4 · in 0.7.3 · in 0.7.2 · in 0.7.1 · in 0.7 · in 0.6.11
A normal day with kwant
Three things you'll see in the app most days.
Keeps going while you sleep
Say "every morning" and the bot sets up the routine. You come back to what changed.
Only risky steps ask first
Ordinary work just runs. Something risky, like a push, a network call or a change outside the bot's folder, shows you exactly what will happen first. You can switch that off too.
Spaces & accounts
Keep work, side projects and home apart
Each space has its own bots, chats, routines and accounts. Nothing leaks between them.
A space for each part of your life
Give it a name, an icon and a colour. Switch with ⌘/Ctrl + 1–9, and each space opens where you left it.
How spaces work →Every account, side by side
Claude, Codex or Grok logins, API keys and your own server. When one runs out of usage, the turn moves to the next and the chat says so.
Accounts & profiles →Your team
A real team, not another chat tab
Each bot has its own role, voice and model, and knows who else is in the room.
Handoffs
Designer answers, then passes the check to QA with an @mention.
Team discussions
The lead runs a short discussion and posts one summary. Open the fold to read every turn.
Approvals
Allow once, always allow or deny, with the full arguments in front of you.
New teammates
Pick a preset like Security, Research or Chef. Name, role, avatar and voice fill in.
How it compares
Other AI tools say "done". kwant shows it.
Chat apps and agent frameworks can do the work. kwant is built to prove it: checked results, quoted sources, and nothing paid without you.
| kwant | Chat assistant apps | Agent frameworks | |
|---|---|---|---|
| "Done" means it was checked | Yes (a card turns Verified only when kwant runs the check itself) | No (you take the model's word for it) | You build it (write your own evals) |
| Facts open at the exact quote | Yes (no quote on the page, no claim) | Partly (a link to the page, rarely the words) | You build it |
| Conflicting facts wait for you | Yes (Truth Ledger shows both; you keep one) | No (one answer, picked for you) | You build it |
| Payments always ask | Yes (even behind a plain "Continue") | Varies (depends on the app and mode) | You build it |
| Keeps working while you're away | Yes (routines, follow-ups and an Inbox of what happened) | Partly (some have scheduled tasks) | Partly (you host and schedule it) |
| Autonomy you can dial up | Yes (Flight School: bots earn tiers per kind of work, only you grant them) | No (one setting for everything) | You build it |
| The right bot answers, and says why | Yes (picks the bot, model and effort) | No (one assistant per chat) | You build it (you write the routing) |
| Any model, with failover | Yes (Claude, Codex or Grok logins you already pay for, API keys, Ollama or any OpenAI-compatible server) | No (their own models) | Yes (you wire up each provider) |
| Separate spaces for work and home | Yes (each with its own bots, accounts, routines and approval rules) | Partly (projects, under one account) | You build it |
| Light to run on your Mac, Linux or server | Yes (one binary with the app and database built in) | No (the provider's cloud) | Varies (you assemble the runtime, queue and database) |
| No code needed | Yes | Yes | No (it's a library you code against) |
| Price | Free and open source; you bring the model accounts | Free tier or a monthly plan | Free library, plus your hosting and models |
OpenClaw vs Hermes Agent vs kwant · kwant vs Grok bot · Self-hosted AI agents
While you're away
It keeps going when you close the tab
Routines run on a schedule, bots check back on their own, and kwant keeps track of what's still open.
Say "every morning", get a routine
The bot sets it up itself, with Undo, Edit and Pause on the card. Or make one: daily, weekly, every few minutes, or a cron rule.
Routines & bot actions →Inbox
What happened while you were away: what's waiting on you, what finished and what got stuck, on one page.
Loose ends
Finds the reply you owe in chats, email and GitHub, and drafts it. It never sends.
Letters
Photograph a bill or notice. kwant pulls out the deadline and amounts and reminds you.
New in 0.6.9
You'll know when a bot needs you
A pill slides in when a bot in another chat waits on you, and the tab title counts what's waiting. When a bot needs a sign-in, you take over its browser.
A pill, wherever you are
An approval, a question or a sign-in slides in at the top for a few seconds. Click it to go there. On a phone too, and under the Mac app's menu-bar icon.
What's new in 0.6.9 →Take over to sign in
The bot posts a sign-in card. Open the page, sign in yourself, and Done hands the browser back. Bots never ask for a password in chat.
Sign-in & take over →Sign-in card
One card per site. Skip, and the bot says what it couldn't do.
On your phone
The same pill at the top of the screen.
What it's doing
A line above the message box says what the bot is using right now.
New in 0.6.11
Answer with a tap, reply, solve the human check
Every finished bot message has 👍, 👎 and Reply. When a bot asks you something, the thumbs answer it. When a site asks if you're human, the bot hands the check to you.
Yes, go ahead
When the bot's newest message asks you something, 👍 means Yes, go ahead and 👎 means No; hover a thumb to see which. Pick one and the bot carries on without you typing. Anywhere else, a 👍 or 👎 is feedback the bot reads next time.
What's new in 0.6.11 →Reply
Reply quotes the message in the box, so the bot knows which one you mean.
Human checks come to you
A CAPTCHA stops the bot and a card in that chat asks you to solve it. Once it's solved, the bot carries on by itself; kwant never solves one itself.
On your phone
The same thumbs and Reply under each answer.
More in kwant
- Parallel UniverseLoad-test up to three ways to deploy on your own server before you pick one.
- Bots that grow with youChanges with an Undo, search across every chat, imports and @mention handoffs.
- From your phoneRead the team's summary or approve an action away from your desk.
- Works with what you haveClaude Code, Codex, Grok, API keys and any MCP server. Mac, Linux, Docker.
Up and running in three steps
You need Rust and Node.js to build it, and one AI account.
-
1
Build and run
make runThe web app opens on your machine. On a server, open the link it prints once.
-
2
Add an account
Claude CodeCodexGrokAPI key
Use a CLI you're already signed in to, or paste a key. Add as many as you like.
-
3
Ask your team
"Every weekday at 9, check the deploys."
The bot answers, or sets up a routine with an Undo button.
Give it a try tonight
One command to start. Your bots, your accounts, your machine.
git clone https://github.com/alexsergivan/kwant && cd kwant && make rungit clone https://github.com/alexsergivan/kwant && cd kwant && docker compose up -d --build
Read the quickstart
Run it on a server
Download for Mac
Download for Windows beta
Download for Linux beta
The Mac app isn't notarized by Apple yet, so macOS stops it the first time: how to open it. Also for Windows and Linux, in beta.
The installer isn't signed yet, so Windows warns the first time: click More info, then Run anyway. Also for Mac and Linux.
A .deb for Ubuntu and Debian, or the AppImage for other distros (how to install). Also for Mac and Windows.
kwant runs on coffee and GitHub stars. Like it? Toss one our way.
Star on GitHub