Kwant 0.8: Self-Hosted AI Bot Team with Enhanced Verification and Security

What changed

Kwant version 0.8 introduces significant enhancements to its self-hosted AI bot team framework, focusing on improved verification, security, and user experience. A key new feature is the ability for bots to visually demonstrate their work; they can now share screenshots directly in chats to show changes, such as misaligned buttons, with before-and-after comparisons. This visual evidence is presented once per turn. For tasks requiring verification, Kwant now runs checks itself, such as tests or file integrity checks, before marking a card as 'Verified' on its Mission board, moving beyond simple bot assertions.

The 'approvals off' setting has been expanded to cover all engines, including Claude, Codex, and Grok CLI, allowing bots to proceed with tasks without explicit user approval for many actions, though sensitive operations like sign-ins, key management, or irreversible changes still require explicit consent. Users can now drag and drop files directly onto any chat interface in the web, Mac, Linux, and Windows apps, with Kwant indicating which bot will receive the file and any transmission limitations.

Importing bot workspaces has been broadened to support formats from OpenClaw, Hermes Agent, Codex's AGENTS.md, and Paperclip. The 'Found on this computer' feature lists local bots by name and count, while a 'Folder' import packs only necessary files, excluding sensitive credentials. GitHub repositories can also be imported via a link, with license information displayed first. Imported routines are paused by default, and previews highlight any omitted components, with secrets never being imported.

Development and code review processes are strengthened with 'Dev mode' reviews and the Swarm Auditor, which operate without chat history or reviewer memories. A 'pre-push card' now inspects code before git push or GitHub file writes, checking for secrets, new packages, and weakened tests, requiring explicit user override with a stated reason for any exceptions. Security on macOS has been tightened, with Claude tools denied access to sensitive folders like sign-in directories, the Keychain, and Kwant's vault when running locally; KWNT_MAC_SEATBELT=1 enforces a stricter sandbox. Other improvements include bots using the Enter key to submit forms, a functional browser for Windows and Linux ARM, and resolved issues with bot deletion.

Why it matters for builders

Kwant 0.8 offers builders a more robust and trustworthy platform for deploying AI agent teams. The enhanced verification mechanisms provide concrete proof of task completion, reducing reliance on bot self-reporting and increasing confidence in automated workflows. The expanded import capabilities and improved security, particularly the sandboxing on macOS and pre-push checks, allow for safer integration of AI agents into development pipelines and sensitive operational environments. This release democratizes advanced agent orchestration with a self-hosted, open-source solution.

Practical impact

Builders can leverage Kwant 0.8 to create more autonomous and reliable AI teams. The visual evidence provided by bots through screenshots can significantly aid in debugging and understanding bot actions. The stricter approval flows for sensitive operations, combined with the ability to customize bot permissions and access, offer a granular control over AI agent behavior, crucial for enterprise adoption. Developers can explore integrating Kwant into their CI/CD pipelines for automated code reviews and testing, benefiting from the pre-push checks. The ability to drag and drop files directly into chats simplifies data input for bots, streamlining workflows. For those concerned about data privacy, the self-hosted nature and lack of telemetry are significant advantages.

Caveats and source limits

The source material details the features of Kwant 0.8 extensively, including its technical implementation in Rust and React, and its operational capabilities. However, specific performance benchmarks or independent comparisons against alternatives like Grok Bot, Claude Duet, or OpenAI dots are not provided. Pricing information is absent, consistent with its open-source, self-hosted model, but details on potential hosting costs or resource requirements for server deployments are not elaborated upon. The beta status for the Windows and Linux x86_64 (AppImage) versions is noted, suggesting potential for further refinement in these platforms. The source also mentions a 'Laya router' which is downloaded on first start, with Python packages and models from PyPI and Hugging Face, but does not detail its specific function or any associated dependencies beyond its availability via --no-laya flag.