Whiteboard: An Open-Source Desktop App for Agent-Assisted Software Architecture Review

Whiteboard is an open-source desktop application designed to facilitate collaborative software architecture between human developers and AI agents. It provides a shared workspace where agents can draw diagrams and explain their work, while humans navigate the resulting code and architectural decisions.

Developed by Sid, Alex, Ketan, and Milan, the tool emerged from the founders' struggle to maintain codebase comprehensibility when using agentic coding. They aimed to recreate the clarity of a traditional whiteboard session, where a deep understanding of a system is achieved, even as AI agents produce large volumes of code.

Built on CodeOSS, the foundation of VSCode, Whiteboard integrates architectural visualizations like sequence diagrams and ERDs directly with the underlying code. Clicking a diagram element or an agent's trace allows users to jump to the relevant code, leveraging familiar VSCode keybindings and Language Server Protocol (LSP) support. This addresses the common challenge of connecting high-level design to implementation details.

A custom semantic, AST-aware diff viewer, written in Rust, filters code changes to show only relevant modifications, summarizing large additions and collapsing boilerplate like unit tests. This aims to reduce what the founders call 'cognitive debt' from reviewing extensive agent-generated code. Additionally, a 'Decision Log' allows agents to query and link their internal traces, offering transparency into their autonomous decision-making process.

"We were missing the feeling of a “whiteboard session” with another dev where you leave with a deep understanding of a system"

Currently, Whiteboard is used at companies like Salesforce and Modal for reviewing architecture or spec-level changes. It helps developers iterate on AI-generated prototypes by making their output more reviewable, and can escalate complex changes from automated code reviewers for human judgment. The desktop app is available for MacOS and Linux under an MIT license, with a hosted web version planned for companies, offering features like trajectory storage and multiplayer reviews.

For a developer working with AI agents, Whiteboard aims to make the output of these tools less opaque and more manageable. It shifts the interaction from merely accepting or rejecting agent-generated code to a more iterative, collaborative design process where architectural decisions and their code implications are clearly linked and reviewable. This moves beyond basic code generation to a more integrated "co-architecting" experience.