Watch it work 0:52

Watch kwant work

kwant demo, 52 seconds. A team of bots picks up a task and hands it on, you point at the exact spot on a page that's wrong, a bot asks before running a risky command, each task shows the checks it passed, and a routine keeps watching overnight. Recorded on a sample demo workspace, with no real account or repo.

Open source·Self-hosted·One Rust binary

Your own
team of bots.

kwant is a self-hosted team of AI agents. They share chats, hand work to each other and keep going on a schedule, on your laptop or a VPS.

It isn't signed yet, so Windows warns the first time: click More info, then Run anyway.

A .deb for Ubuntu and Debian on x86_64; other distros, the AppImage (how to install).

A VPS or Docker? Get started On your phone? No app store needed
First time on a Mac? How to open it

macOS asks once, because kwant isn't notarized by Apple yet.

  1. Drag kwant to Applications and open it. When macOS says it couldn't verify it, click Done.
  2. Open System Settings › Privacy & Security and scroll down to Security.
  3. Click Open Anyway next to “kwant” was blocked, then Open Anyway again, and enter your Mac password or use Touch ID.
A mock-up of the three macOS steps: the “kwant” Not Opened alert with Done marked, System Settings › Privacy & Security, and the Security row “kwant” was blocked to protect your Mac with Open Anyway marked.

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.

0.9.18 released Nov 16 54 releases since Oct 25 AGPL-3.0, open source

Designer answers, then hands the check to QA with one @mention.

Full size
Designer answers, then hands the check to QA with one @mention.

A normal day

They work while you don't.

Give each bot a job and a schedule. You come back to finished work, and to the one thing that needs you.

  1. Research runs its morning routine and posts three competitor changes in #market.

  2. TechLead passes the pricing page to Designer with one @mention. Designer picks it up.

  3. QA wants to run gh pr create --base main and asks first. Allow once

  4. Inbox: 4 done, 1 waiting on you. Each done task lists the checks it passed. 4 checked

The Inbox since you last looked. Waiting on you: Ann's high-risk task Speed up the search box, marked Needs you with medium confidence, and her question about the pack size guide with cm, inches and Both buttons. Nothing failed or finished, and one task waits in To review. On the right, the search task's verdict says it couldn't time the full catalogue, with Approve, Send back and Roll back.
The Inbox at 18:05. The amber row is the one that waits for you.

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 →

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 →

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 →
The space rail with three spaces, Trailhead, Pixel Quest and Home, each with its own icon and colour. Trailhead is open on #general with a weekly digest posted by a routine.

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 →
Settings, This space, Apps and accounts, AI accounts: Claude is used first, ready and used by Ann and Ben. If it runs out: Codex, ready, which Cody uses first, a Personal Max Claude login out until 06:34, and Ollama on the Mac, which can't connect, each with Use first. Grok isn't used and isn't installed. Below: Sign-ins and settings, and Advanced.

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.

The Product team chat. A question about adding a trip waitlist, a collapsed "Team discussed, 4 turns" fold, and TechLead's summary covering flow, backend, hold length, UI and the release gate.

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.

Kinds of tool, not named products. Each one varies; check its own docs. Flight School and the Truth Ledger are opt-in.
kwantChat assistant appsAgent 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)
PriceFree and open source; you bring the model accountsFree tier or a monthly planFree library, plus your hosting and models

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 →
The Inbox since you last looked. Waiting on you: Ann's high-risk task Speed up the search box, marked Needs you with medium confidence, and her question about the pack size guide with cm, inches and Both buttons. Nothing failed or finished, and one task waits in To review. On the right, the search task's verdict says it couldn't time the full catalogue, with Approve, Send back and Roll back.

Inbox

What happened while you were away: what's waiting on you, what finished and what got stuck, on one page.

Loose ends. Designer has waited four days for your pick of a hero image, and you promised Mara a press kit by email. Each case quotes the message it is based on and has a draft reply.

Loose ends

Finds the reply you owe in chats, email and GitHub, and drafts it. It never sends.

A scanned letter from a fictional water company next to what kwant read from it: who it's from, what they want and the deadline of November 5, each with the quote it came from and editable.

Letters

Photograph a bill or notice. kwant pulls out the deadline and amounts and reminds you.

Loose ends & letters →

Up and running in three steps

You need Rust and Node.js to build it, and one AI account.

  1. 1

    Build and run

    make run

    The web app opens on your machine. On a server, open the link it prints once.

  2. 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. 3

    Ask your team

    "Every weekday at 9, check the deploys."

    The bot answers, or sets up a routine with an Undo button.

Read the quickstart →

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 run

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.