kwant

Product Introduction

1. Definition: kwant (pronounced "kwnt") is an open-source, self-hosted AI agent team platform, packaged as a single Rust binary. It falls into the technical categories of autonomous AI agents, multi-agent systems, and self-hosted AI orchestration frameworks.
2. Core Value Proposition: It exists to provide a private, verifiable, and collaborative AI workforce that runs on your own infrastructure (Mac, Linux, or VPS). Its core value is moving beyond single-chat AI assistants to a team of specialized bots that autonomously collaborate, verify their own work, and operate on a schedule, all while giving the user granular control and transparency.

Main Features

1. Multi-Agent Team with Intelligent Routing: kwant deploys a team of up to 27 specialized AI agents (bots) with distinct roles (e.g., TechLead, Designer, QA, Research). It uses an intelligent routing system where each user message is automatically directed to the most suitable bot based on its defined "Good for" and "Not for" rules. The selected bot explains why it was chosen, and misrouted tasks can be manually handed off with one click, teaching the system.
2. Verified Task Completion & Truth Ledger: Unlike standard AI chats that output unverified text, kwant implements a verification system. Bots break work into cards and hand them off for peer testing. A task card only turns "Verified" after kwant itself executes a check. The optional "Truth Ledger" feature flags conflicting facts (e.g., from web research) and holds both until the user manually adjudicates, ensuring a single source of truth.
3. Controlled Autonomy with Flight School & Approvals: Bots operate with configurable autonomy. The "Flight School" system allows bots to earn higher autonomy tiers for specific types of work, but the user must explicitly grant each level. Crucially, any action involving a payment, network call, or change outside its sandboxed folder triggers a mandatory approval screen showing the exact command or action, with options to Allow Once, Always Allow, or Deny.
4. Isolated Workspaces (Spaces) & Multi-Model Failover: Users can create separate "Spaces" (e.g., for work, personal, side projects), each with isolated bots, chats, routines, and AI accounts. kwant supports multiple AI backends simultaneously, including direct logins to Claude (via Claude Code), Codex, and Grok, as well as API keys and OpenAI-compatible servers (like Ollama). It provides automatic failover, switching to the next available account/model if one hits a usage limit.
5. Persistent Operation & Routine Automation: The system is designed for continuous operation. Bots can be assigned routines (e.g., "every weekday at 7 AM") that run on schedule even when the web interface is closed. Features like "While You Were Away" inbox, "Loose Ends" (drafting pending replies), and "Letters" (extracting data from document photos) enable asynchronous productivity.

Problems Solved

1. Pain Point: The lack of verification and accountability in standard AI chat interfaces, where the user must trust the model's output without independent validation.
2. Pain Point: The inefficiency and context loss of manually copying prompts between different single-purpose AI assistants or chat tabs.
3. Pain Point: The risk and opacity associated with granting AI agents the ability to perform actions (like git commits, API calls, or payments) without explicit, granular consent.
4. Target Audience: Developers, engineers, and technical founders seeking to automate workflows (code review, deployment checks, research). Solopreneurs and small teams needing a scalable, automated "assistant team." Privacy-conscious users and enterprises requiring full data control by self-hosting AI operations.
5. Use Cases: Automated daily stand-up reports and competitor monitoring. Managing and triaging GitHub issues or pull requests. Breaking down a complex product feature request into design, development, and QA tasks handled by specialized bots. Automating weekly personal finance review from bills and statements.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Unparalleled transparency and control for an agentic system, with mandatory approvals and verifiable results. True multi-agent collaboration with handoffs. Lightweight deployment via a single binary. Privacy-focused and cost-effective as it uses existing AI subscriptions (Claude, Codex, Grok). Open-source (AGPL-3.0) allows for audit and customization.
   • Cons: Requires technical comfort to self-host and maintain. Relies on the user providing and funding their own AI model accounts. The macOS binary is not notarized, requiring manual security override on first run. As a younger project, it has a smaller ecosystem and community compared to established frameworks.

2. Key Alternatives & Differentiation:

   • vs. Chat Assistant Apps (Claude.ai, ChatGPT): These are general-purpose chat interfaces with single, non-specialized assistants. They lack kwant's multi-agent collaboration, scheduled automation, verification systems, self-hosting, and granular action approvals. They operate in the vendor's cloud, not on your hardware.
   • vs. Agent Frameworks (AutoGPT, LangChain): These are code libraries for developers to build custom agent systems. kwant differentiates by being a ready-to-use, no-code (or low-code) end-user application with a built-in UI, pre-configured agent roles, and integrated features like the Truth Ledger and Flight School. It bundles the runtime, queue, and database into one binary, whereas frameworks require assembly.
   • vs. Other Self-Hosted Agents: Many solutions are Docker-heavy or complex to deploy. kwant's single Rust binary offers a significantly simpler deployment story for Mac and Linux, with a focus on user experience, team dynamics, and built-in safety controls.

Frequently Asked Questions (FAQ)

1. Is kwant free to use? Yes, the kwant software itself is completely free and open-source under the AGPL-3.0 license. However, you must provide and pay for your own AI model accounts (e.g., Claude, Codex, Grok subscriptions or API keys) for the bots to function.
2. What are the system requirements for running kwant? kwant can run on any modern Mac (Apple silicon or Intel) or Linux (x86_64) system, including a personal laptop or a cloud VPS. It is distributed as a single binary or via Docker. You need an active account with an AI model provider for it to connect to.
3. How does kwant ensure safety when bots perform actions? kwant employs a mandatory approval system. Any action deemed "risky" (file writes outside its folder, network calls, git commands, etc.) triggers a prompt showing the exact action. The user must choose "Allow Once," "Always Allow," or "Deny." This is complemented by the "Flight School" where autonomy is granted per task type.
4. Can I customize the AI agents or create new ones? Yes. You can modify existing bot roles, voices, and instructions. You can also create new bots from presets (like Security, Research) or from scratch, defining their purpose, model preference, and the websites they are allowed to browse.
5. Does kwant work with local LLMs like Llama or Mistral? Yes, through compatibility layers. kwant can connect to any OpenAI-compatible API server. This means you can use it with local models served by tools like Ollama, LM Studio, or text-generation-webui by configuring them to expose an OpenAI-compatible endpoint.