Whiteboard, a YC-backed open-source IDE, lets coding agents draw their own architecture diagrams

A team of four college friends has released Whiteboard, an open-source desktop IDE that lets developers and AI coding agents sketch software architecture together on a shared canvas. The app, built on CodeOSS, the open-source core of VS Code, plugs into agents including Claude Code and Codex and gives them an SDK for drawing sequence diagrams, entity-relationship diagrams and trace excerpts directly into the workspace. It ships under an MIT license for macOS and Linux, with source on GitHub, and launched under the Y Combinator W26 batch.

The founders are Sid, Alex, Ketan and Milan, who describe quitting tech-lead jobs just as AI coding assistants became standard. Their argument is that tradeoffs in agent-written code often surface only after a first implementation pass, and that pull requests started merging faster than anyone could read them. They call the resulting drift cognitive debt, a term they credit to Geoffrey Litt, and pitch design review as a way to catch it before the codebase becomes hard to contribute to.

Whiteboard's core interface is deliberately different from a text diff. Clicking a diagram or a quoted line from an agent's trace jumps straight to the underlying source, and developers keep VS Code keybindings and language-server support without configuration. The company says engineers at Salesforce and Modal already use the tool to review architecture and spec-level changes, an early signal that the workflow has traction beyond the founders' own projects.

Two components carry much of the technical load. A semantic diff viewer written in Rust parses the AST so large added functions collapse into pseudocode summaries, while unit tests and bulky documentation edits are folded away by default. A WASM-based plugin system lets teams override those defaults, and a separate Decision Log gives agents a way to query and link their own traces back to the canvas, so a reviewer can see how a stated requirement turned into code and which choices the agent made independently.

The founders position Whiteboard as complementary to automated reviewers like Greptile, suggesting teams run a bot on small changes and escalate only the ones needing human judgment to a Whiteboard session. That split, machine triage followed by human design review, puts the product in the same conversation as code-review startups and IDE vendors racing to add agent features. The business model is a hosted web version with session management, trajectory storage and multiplayer review sold to companies, while the desktop app and self-hosting stay free.

The open-core bet raises a familiar question: whether teams accustomed to owning their code and traces will pay for hosted convenience or keep everything on their own machines. An earlier HTML-based prototype hit a wall before the founders moved to CodeOSS, and the repo is public as they ask for contributions and feedback. What to watch is whether the semantic diff and Decision Log become standard parts of agent review workflows, or whether IDE vendors absorb those features first.