Whiteboard is an open-source desktop app that gives coding agents — Claude Code, Codex, and others — a shared visual workspace to explain their work to human developers. The core thesis, drawn explicitly from Geoffrey Litt's 2026 essay, is that understanding has replaced writing as the primary bottleneck in software engineering. When agents can generate thousands of lines in minutes, the human's job shifts from authorship to review, and existing tools were not built for that job. The product is a vendored fork of Code OSS (Microsoft's open-source VS Code base) stripped of roughly 45% of the original codebase — mostly Copilot-related code — and rebuilt around three features: diagrams that link to code, an AST-aware semantic diff viewer written in Rust, and a decision log that lets agents query and visualize their own reasoning traces. The semantic diff is the most technically interesting piece: it collapses noise like test additions and documentation changes, summarizes large new functions as pseudocode, and is extensible via a WASM plugin system. This is not a chatbot wrapper. It is infrastructure for a specific workflow. The architectural choice to vendor rather than patch Code OSS is revealing. The team argues that patch-based forks are hard for coding agents to maintain — an admission that agents are already part of their own development workflow. They monitor upstream for security and feature patches and merge selectively. The result is a full desktop editor that inherits VSCode's LSP support and keybindings but is purpose-built for review rather than authoring. The project is MIT-licensed, runs entirely against local checkouts, and collects only anonymous telemetry that explicitly excludes code, diffs, prompts, and model output. A hosted team product is planned, but self-hosting will remain available. This is the familiar open-core playbook: build trust and adoption with a generous open-source base, then monetize collaboration features for teams. The known limitations are honest and telling. You cannot edit files in Whiteboard — it is deliberately read-only for now. Multi-repo reviews are not well supported. Shared reviews are snapshots, not live documents. These constraints reinforce the thesis: this tool is for understanding, not authoring. Whether that constraint is principled design or a shipping-speed tradeoff remains to be seen. The recommended models — GPT-6 Sol and Claude Opus 5.5 — place this firmly in the frontier-model tier, which narrows the immediate audience to teams already paying for top-tier API access. The value proposition only works if the agent's output is complex enough to need visual explanation, which means large-scale codebases and multi-commit changes. For solo developers writing small features, the overhead of a dedicated review app may not clear the bar. The deeper question is whether this category — agent-to-human explanation tooling — becomes as essential as the IDE itself. If AI-generated code volumes continue to grow, the review bottleneck will only sharpen. Whiteboard is early to this bet, open-source, and technically serious. The risk is that the major IDE vendors (Microsoft, JetBrains) absorb the semantic-diff and decision-log features into their own products before Whiteboard builds a defensible user base.
