kwant
A self-hosted team of AI bots that hand off work, ask before risky steps, and count a task done only when kwant runs the check itself.
- Source repo
- alexsergivan/kwant
- Stars
- ★ 681
- Last updated
- 1d ago
- License
- AGPL-3.0
- Primary language
- Rust
- FA score
- 68/100 · Some gaps
At a glance
- Works with
- Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
- You'll need
- Typical use
- A small product team wants a role-separated bot crew (architect, backend, QA, marketer) on its own server, handing tasks back and forth in channels instead of isolated chat tabs.
- Main limitation
- First start downloads Laya and a ~420 MB model from PyPI and Hugging Face; Laya idles at roughly 1.7–2.3 GB and the full Compose stack wants about 4 GB free RAM.
- Website
- kwant.si ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release mac-v0.9.32 (2026-11-20)+24.0 ★/day over 7 days
- Source review
- 68/100 · Some gaps
What does this agent do, and when should you use it?
kwant is a self-hosted multi-bot platform written in Rust (Tokio, Axum, SQLite) with a React web app embedded in the binary, shipped as a macOS/Linux/Windows (beta) desktop app, a single kwnt binary, or a Docker Compose stack. Inside separate spaces it gives you bots with their own roles, voices and models, channels, teams, routines and accounts; bots @mention each other to hand work on, and kwant picks which bot answers each channel message and says why. Work lives as cards on a Mission board, and a card turns Verified only when kwant runs the test or file check itself and keeps the receipt. It reuses the Claude, Codex or Grok CLI logins you already have, or Anthropic/OpenAI/xAI API keys, with per-engine accounts, a priority order and automatic failover when usage runs out. Approvals default to asking only for risky steps, payments always ask, and Flight School lets each bot earn more freedom per kind of work.
In a channel, kwant reads each incoming message, applies routing rules plus the Laya model router to pick a bot, and shows the reason; the chosen bot answers using its role, soul and model through the configured engine (Claude/Codex/Grok CLI, an Anthropic/OpenAI/xAI key, or a self-hosted OpenAI/Anthropic-compatible server). Bots hand work on with @mentions, and a team lead can run short discussions and post one summary. Work becomes Mission board cards: on review, a teammate on another model re-runs the checks in a clean copy of the files, perturbs the code for high-risk work, and reports Pass only when kwant re-ran the check itself. In Dev mode the Dev desk shows what landed in a repo in plain words, a bot that did not write the PR reviews it (tests and secrets included), failed CI checks and review comments return to the bot as one Fix-it item, and nothing is ever merged. The approval layer intercepts git push, network calls, unknown commands and changes outside the bot's folder with the exact command and Allow once / Always allow / Deny; payment commands and browser checkouts always ask. Routines run on a schedule (bots can create them on request, with Undo), the Inbox groups what needs you, what finished and what failed, and per-task git worktrees, Watch-live Point selection, diff-line notes, Truth Ledger quote verification against published pages, Loose ends from chats and read-only IMAP/GitHub, and Parallel Universe deployment comparison with local containers are also built in. All state stays in ~/.kwnt (or $KWNT_DATA_DIR) with no telemetry.
- A small product team wants a role-separated bot crew (architect, backend, QA, marketer) on its own server, handing tasks back and forth in channels instead of isolated chat tabs.
- A developer wants bots writing code with approval cards before push, CI failures routed back to the authoring bot, and a fresh-eyes review from a different model.
- A solo user with Claude, Codex and Grok accounts wants priority-ordered failover so work continues when one account hits its usage limit.
- Someone who needs overnight or scheduled routines (e.g. check checkout errors every morning) and an Inbox that shows results and failure reasons.
- A team doing work that touches payments, where every checkout must stop for human approval regardless of the approval mode.
- Users migrating existing bot personas from a Claude Code subagent, a SKILL.md, a character card or a JSON export.
How do you install or deploy this agent?
The quickest path is the desktop app: download the universal macOS .dmg, the Linux .deb or AppImage, or the Windows x64 beta installer from Releases. You need at least one AI account (a Claude/ChatGPT/Grok login or an Anthropic/OpenAI/xAI API key); the Get started card installs a missing CLI for you.
Building from source needs Rust stable and Node.js 20+:
git clone https://github.com/alexsergivan/kwant.git
cd kwant
make run # builds frontend/dist, then runs cargo run -- serveFor a release binary:
make build
./target/release/kwnt serveOn a VPS with Docker Compose (needs Compose and about 4 GB free RAM):
KWNT_CLIS=claude,codex docker compose up -d --build
docker compose exec kwnt kwnt doctor
docker compose exec kwnt claude auth login
docker compose exec kwnt kwnt signin-linkWithout Docker, kwnt setup installs the browser, Laya and a CLI, and kwnt doctor reports what is missing. First start installs Laya and its roughly 420 MB model in the background; pass --no-laya to skip it.
How do you use this agent?
Open http://127.0.0.1:8799 after starting; a new #general shows the Get started card where you choose an AI, sign in and pick a team. A fresh install ships Architect, TechLead and DevOps. On a server, kwnt serve --host 0.0.0.0 --port 8799 prints a one-time sign-in link valid for 15 minutes, or use:
kwnt signin-link
kwnt token --rawAdd several accounts and set the default and priority order under Settings → This space → Accounts. Ask in a channel and kwant picks the fitting bot with a reason; Wrong bot? corrects it next time. Track work on the Mission board and handle review items in the Inbox with Approve, Send back or Roll back. Asking a bot for “every morning at 08:30” creates a Routine with Undo. Approval cards appear in chat with Allow once / Always allow / Deny, and payments always ask. For phone notifications, install the web app and enable Web Push under Settings → All spaces → Notifications.
What are this agent's strengths and limitations?
- A concrete definition of done: Mission board cards become Verified only after kwant re-runs the test or file check itself, with a receipt, and attachments alone never count as proof.
- Genuine multi-engine, multi-account support: reuse Claude, Codex or Grok CLI logins or Anthropic/OpenAI/xAI keys, set a priority order, fail over when usage runs out, and pin a bot to one account.
- Layered approvals: Ask only when risky by default, payments always ask in every mode, and Flight School grants more autonomy per kind of work only when you allow it.
- Three real deployment shapes: desktop app (macOS/Linux/Windows beta), one self-contained kwnt binary, or Docker Compose, with SQLite and the web UI embedded and data kept in ~/.kwnt.
- No telemetry and no third-party interface loads; apart from the first Laya download, traffic goes only to the providers you configure.
- Dev mode ships worktrees, PR review with secrets and test checks, CI feedback back to the bot and diff-line notes, while never auto-merging.
- First start downloads Laya and a ~420 MB model from PyPI and Hugging Face; Laya idles at roughly 1.7–2.3 GB and the full Compose stack wants about 4 GB free RAM.
- Inside the Compose stack kwant's own execute_shell tool is unavailable because it would need the host Docker socket, leaving only the CLIs' own shell tools.
- Windows is beta: no self-update, and Grok, VM mode and looking inside videos are unavailable; the Linux desktop build has only been tested in containers.
- No ARM Linux app, and source builds must run make frontend first or cargo build and cargo test fail.
- AGPL-3.0: serving modified kwant over a network obliges you to publish your changes, and a closed-source commercial use needs a separate licence.
- The desktop apps are ad-hoc signed and not notarized, so macOS and Windows both require bypassing an OS warning on first launch.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| kwant This agent | 68 · Some gaps | — | ★ 681 | 1d ago | Rust | Codex · Claude Code · OpenAI API · Claude API |
| DashClaw | 74 · Some gaps | CLIFree | ★ 320 | 7d ago | TypeScript | Codex · Claude Code |
| Jelly: Local Multi-Bot Agent Workspace | 53 · Major gaps | — | ★ 67 | today | TypeScript | ChatGPT · OpenAI API |
| JoySafeter | 53 · Major gaps | Self-hosted serviceFree + model costs | ★ 319 | 7mo ago | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
least_privilege: docs/approvals.md states smart mode is an accident guard, not a sandbox; scripts, make, cargo, npm run inside the bot folder without a card and can reach the network and kwnt's own API. ask mode and path/credential denylists exist, but the default smart mode is broad, so deducted. user_confirmation: approval cards, payments always ask, per-space/per-bot overrides, and bot tools refusing approval changes are well evidenced; full marks. data_flow_transparency: README claims no telemetry and only a first-start Laya download, but no auditable egress list or code-level evidence is provided; deducted. sensitive_data_handling: non-dumpable process, Bitwarden crypto, credential path denylist are present, but SECURITY.md admits Chrome DevTools port is unauthenticated and a bot with a shell can read autofilled passwords; known unfixed risk, deducted. dependency_security: Cargo.toml uses loose version ranges and no Cargo.lock or dependency audit/vulnerability scan evidence is in the packet; deducted. external_effects: payments and production steps always ask, pushes require confirmation, but smart mode scripts can cause external side effects; deducted. rollback: rollback, worktree removal preserving branches, and approval rollback demotion exist, but no general data/state rollback mechanism is shown; deducted. source_attribution: README and Cargo.toml name the repository, homepage and AGPL-3.0, but publisher identity is unverified; deducted.
self_consistency: README, docs/approvals.md, SECURITY.md and source broadly agree on approval modes, payments and known limits, but README's 'kwant has no token' sits awkwardly beside the documented API token; deducted. dependency_availability: dependencies are common crates.io libraries and CI uses --locked, but no Cargo.lock is in the packet, so locked versions cannot be statically confirmed; deducted. failure_messages: main.rs doctor/token errors are specific and print fix commands, but the sample is limited; deducted.
audience_and_scenarios: README covers individuals, teams, VPS, Docker, multiple spaces, accounts and routines with strong evidence; full marks. capability_boundaries: Good for/Not for, browser site rules and approval modes exist, but smart-mode boundaries are broad and heuristic; deducted. trigger_precision: @mention routing, Laya routing and routine scheduling are described, but no code evidence of trigger precision or misfire handling is in the packet; deducted. environment_fit: supports macOS/Linux/Windows beta, Docker and VPS, but Windows lacks features and Linux is untested on real desktops; deducted.
information_architecture: README is well structured and docs are topic-split, but the packet lacks a full directory index; deducted. install_notes: installation is thorough, covering macOS signing bypass, Linux deb/AppImage, Windows SmartScreen, VPS, Docker and systemd; full marks. naming_stability: CLI and API naming is consistent, but the project is pre-1.0 and mixes kwant/kwnt; deducted. examples_and_faq: quickstart and screenshots exist, but no standalone FAQ; deducted. known_limitations: SECURITY.md and README explicitly list known limitations and beta gaps; full marks. license: LICENSE is the full AGPL-3.0 text and Cargo.toml declares AGPL-3.0-only, consistent; full marks. versioning_changelog: README references CHANGELOG and 0.9 changes, but the CHANGELOG file is not in the packet; deducted. maintenance_responsibility: SECURITY.md gives a reporting channel and supported versions, but the publisher is unverified and no maintainer list or governance is shown; deducted.
output_usability: outputs are chats, mission cards, approval cards and evidence receipts, which are designed to be usable, but static samples cannot verify actual output quality; deducted. marginal_value: multi-bot handoffs, verified receipts, approvals and routines add differentiation, but the space is crowded; deducted. cost_benefit: self-hosting, ~420 MB model download, 4 GB RAM and multi-account setup are non-trivial costs; deducted.
claim_traceability: README claims point to docs and source files, but some claims (e.g. no telemetry) lack code-level evidence; deducted. cross_source_corroboration: README, docs/approvals.md, SECURITY.md and CI workflows corroborate each other, but the packet is incomplete; deducted. fact_inference_separation: docs distinguish verified/estimate/speculation for Parallel Universe, but overall separation of claims from implementation evidence is incomplete; deducted.
- smart approval mode is not a sandbox: scripts, make, cargo and npm run inside the bot folder run without a card and can reach the network and kwnt's API; use ask mode for untrusted bots.
- SECURITY.md states the Chrome DevTools port is unauthenticated and a bot can read autofilled passwords; keep approvals on when using saved passwords.
- Publisher identity is unverified; evaluate installer and update-path trust yourself. The macOS app is ad-hoc signed and the Windows installer is unsigned.
- First start downloads a ~420 MB Laya model from PyPI/Hugging Face; confirm network and disk budget.
- No Cargo.lock is in the packet and dependency versions are loose ranges, so build reproducibility and dependency vulnerability status cannot be statically confirmed.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
mac-v0.9.32 kwant 0.9.32 for Mac and Linux
Release notes excerpt
kwant 0.9.32 for Mac and Linux. Mac One universal app for Apple silicon and Intel Macs (macOS 11 or later). Install: open kwant_universal.dmg and drag kwant to Applications. First open: the app is not notarized by Apple yet, so macOS blocks it the first time. 1. Open kwant from Applications. macOS says it can't check it for malicious software. Click Done (or OK). 2. Open System Settings → Privacy & Security, scroll down to the message about kwant and click Open Anyway. 3. Confirm with Open Anyway and your password. From then on it opens normally. …View on GitHub ↗ -
mac-v0.9.31 kwant 0.9.31 for Mac and Linux
Release notes excerpt
kwant 0.9.31 for Mac and Linux. Mac One universal app for Apple silicon and Intel Macs (macOS 11 or later). Install: open kwant_universal.dmg and drag kwant to Applications. First open: the app is not notarized by Apple yet, so macOS blocks it the first time. 1. Open kwant from Applications. macOS says it can't check it for malicious software. Click Done (or OK). 2. Open System Settings → Privacy & Security, scroll down to the message about kwant and click Open Anyway. 3. Confirm with Open Anyway and your password. From then on it opens normally. …View on GitHub ↗ -
mac-v0.9.30 kwant 0.9.30 for Mac and Linux
Release notes excerpt
kwant 0.9.30 for Mac and Linux. Mac One universal app for Apple silicon and Intel Macs (macOS 11 or later). Install: open kwant_universal.dmg and drag kwant to Applications. First open: the app is not notarized by Apple yet, so macOS blocks it the first time. 1. Open kwant from Applications. macOS says it can't check it for malicious software. Click Done (or OK). 2. Open System Settings → Privacy & Security, scroll down to the message about kwant and click Open Anyway. 3. Confirm with Open Anyway and your password. From then on it opens normally. …View on GitHub ↗
FAQ
Is kwant open source?
Is kwant safe to use?
Is kwant still maintained?
Who makes kwant?
What is kwant's official website?
Thanks — we use this to improve the page.