Whiteboard launches an open-source IDE for AI agents and humans
AI writes the code, but humans still need to understand it. Whiteboard is a new canvas that helps you review agent decisions visually.
Tools · Source: Hacker News
What happened
YC W26 startup devdotfast just dropped Whiteboard on Hacker News. It is an open-source desktop app built specifically for the agent era. Humans and AI models architect software together in a shared workspace. The project already has 1.5k stars on GitHub. It runs entirely on your local machine against your local checkouts.
The tool plugs directly into coding agents you already use, like Claude Code and Codex. It gives your agent an SDK to draw diagrams on an in-app canvas to explain its work. You can click a node in an entity relationship diagram or a sequence diagram and jump straight to the underlying code. The app uses VSCode keybindings and LSP support out of the box so you do not lose your muscle memory.
They built a custom AST-aware diff viewer in Rust to handle agent output. Raw diffs from agents are simply too noisy. Whiteboard summarizes large added functions into pseudocode and collapses boilerplate like unit tests or documentation. It also includes a decision log. Agents can query and link their own traces on the board. This lets you visualize the requirements you set and understand exactly why an agent made an autonomous choice.
Key facts
- 1.5k — GitHub stars at the time of the Hacker News launch
- 45% — Amount of stock VS Code codebase dedicated to Copilot that was removed
- GPT-6 Sol — Recommended model for intelligence, cost, and speed tradeoff
- Claude Opus 5.5 — Recommended model for optimal performance with Whiteboard
Why it matters
The bottleneck in software engineering has shifted from writing code to understanding it. We use dedicated agent TUIs to generate massive pull requests in seconds. Reviewing those line-by-line diffs in a standard text editor no longer works. Whiteboard treats the agent as a junior engineer savant. It gives you the visual tools to manage and review its output rather than just generating more text.
They made a bold architectural choice by vendoring Code OSS instead of maintaining patches. The creators noted that coding agents struggle with patched forks. They stripped out the 45 percent of stock VS Code that is just Copilot bloat. This signals a massive shift away from legacy autocomplete assistants toward dedicated, agent-first review environments. A hosted product for teams is planned, but the core will remain self-hostable and MIT-licensed.
For builders
Build WASM plugins for diffs
Whiteboard's semantic diff viewer is customizable via a WASM-based plugin system. Tooling companies and engineers can build custom views for their specific frameworks to make agent reviews even faster.
Integrate your custom agents
The app provides an SDK for agents to draw directly on the canvas. Founders building specialized AI coders can hook into Whiteboard to give their users a much better visual experience.
Self-host for enterprise privacy
The tool runs against local checkouts and is MIT-licensed. Enterprise teams who cannot send proprietary code to cloud dashboards can run this entirely locally without privacy risks.
My take
We are finally admitting that reading AI generated code is a nightmare. Generating code is a solved problem, but managing the technical debt these agents create is the real billion dollar business. Whiteboard gets it right by focusing entirely on review and comprehension instead of just writing more lines. Stop building faster typers and start building better readers.
Original reporting: Hacker News. This is my rewrite and opinion.