Whiteboard is an open source desktop app for reviewing the code your AI agent writes with sequence diagrams connected to the code, summarized diffs, and a log of the decisions the agent made on its own. It was created by /dev/fast, a YC W26 startup, and works alongside Claude Code and Codex instead of replacing them. It’s MIT licensed and runs on macOS and Linux.
Status at the time of publishing this note (September 25, 2026): Whiteboard is marked as beta, the latest version is v0.1.2 and new versions come out several times a week. Check the repository before following the steps below.
What problem does Whiteboard solve?
It solves the gap between how fast your agent writes code and how fast you can understand it. Its founders are four ex tech leads who hit that wall on their own project. As more PRs written by agents were merged without anyone fully understanding them, they started to feel what they call “cognitive debt”. It got to a point where it was hard for them to contribute to their own code.
If that sounds familiar, it’s the same problem we covered in The New Technical Debt Isn’t in the Code. Whiteboard is a concrete attempt to solve it. The idea is to recreate the session in front of a whiteboard with another developer, where you come out actually understanding the system.
What is Whiteboard?
Whiteboard is a code review environment with an integrated canvas. It’s not another code agent. You keep using Claude Code, Codex, or whatever agent you already have. Whiteboard gives that agent an SDK so it can draw a description of its own work on a canvas within the app.
Under the hood it’s a vendored copy of Code OSS, the open source foundation of VS Code. When you jump from a diagram to the code, you get VS Code’s keyboard shortcuts and LSP support without any configuration. The team vendors Code OSS instead of maintaining patches on top, and says it regularly incorporates upstream security patches.
One important detail: you can’t edit files in Whiteboard, at least in v0.1.2. It’s a place to read and review, not to write.
What does Whiteboard show that a normal diff doesn’t?
It shows three things that a normal diff view doesn’t give you.
Sequence diagrams and entity-relationship diagrams connected to the code
The agent can draw sequence diagrams and entity-relationship (ER) diagrams of the change it made. A sequence diagram shows step by step how components in a flow pass messages to each other. An ER diagram shows data tables and how they relate. Click on a step in the sequence diagram and you go to the exact lines that implement it. The same happens with citations from the agent’s trace, which is the log of what it did during the session.
The founders started with HTML artifacts and abandoned them for this reason. A static diagram is hard to connect to the code, and trade-offs almost always show up only after a first implementation.
A semantic diff instead of a line-by-line one
Whiteboard includes a diff viewer written in Rust that understands the structure of the code, not just lines of text. This is what’s called a semantic diff, and the library is open source: diffr. Right out of the box:
- large new functions are summarized as pseudocode, which you can expand to the real code;
- function signatures always stay visible;
- changes in tests and documentation are collapsed or hidden.
All of this can be customized with WASM plugins.
A log of the agent’s decisions
It’s the decision log: a record of what requirements you defined and what decisions the agent made without asking. Agents can consult their own traces and link them in the whiteboard. This way you can check if an instruction like “use a single transaction for history and resume state” actually made it to the implementation, and spot decisions no one asked for.
How to install Whiteboard and start using it with Claude Code or Codex?
You download the app, connect your agent, and ask it for a review. The README documents three steps:
- Download Whiteboard from install.dev.fast and open the app.
- From the welcome screen, connect Claude Code, Codex, or another code agent.
- Ask your agent to review your current branch against an updated
mainand open the result in Whiteboard.
That’s all the setup the project documents. At the time of publishing this note there’s no separate documentation site, just the README and a Discord server. The release notes also show that OpenCode support arrived in version 0.0.32.
The example prompts in the README suggest a conversational approach. For an API change, tell the agent what your commits are trying to achieve. Then ask it to present the proposed API, examples, and motivation, and only then ask it to explain the implementation.
If something on the canvas doesn’t work for you, select it and copy it. Whiteboard leaves the selection in your clipboard. Paste it in your agent’s chat and ask it to draw it again.
Does Whiteboard work on Windows?
No. At the time of publishing this note, Whiteboard only runs on macOS and Linux, and the only build for Linux is for Fedora. One of the founders said on Hacker News that you can request other distributions. There’s no build for Windows and no announced plan to have one.
Is Whiteboard free?
Yes. The desktop app is MIT licensed, free, and works on your local git copies. The website announces a hosted version for teams as “coming soon”. On Hacker News, one of the founders was less definitive: for now the app is free, open source, and local only, and they’re still deciding about the hosted version. The team says whatever they build will still be self-hostable.
The app collects anonymous telemetry. According to the project, it doesn’t include your code, your diffs, your prompts, or model output, and you can turn it off.
Why not ask the agent for a diagram in Mermaid?
That was the strongest criticism in the Show HN thread. A text diagram that your agent writes to a file costs nothing and doesn’t require a new app. One commenter also argued that Whiteboard is really just a single feature (a server with a graphical interface on top) rather than a complete product.
The difference is the connection to code. A Mermaid file is an image. In Whiteboard, a step in a sequence diagram is a link to a range of lines, with an editor and LSP behind it. Plus you get semantic diff and decision log, which a diagram file doesn’t give you. Whether it’s worth it depends on how much agent-written code you actually review. Also keep in mind that one commenter reported the macOS download is 736 MB.
How does Whiteboard fit with automatic code review?
Whiteboard handles changes that need a person, and an automatic reviewer handles the rest. The founders suggest combining it with an automatic reviewer like Greptile. Small changes are reviewed automatically, and those requiring human judgment scale to a Whiteboard session.
It’s the opposite of tools where one agent reviews another. We covered that approach in How to do code review with Claude Code and other agents with Open Code Review. Whiteboard is for when you need to understand the change yourself.
The founders say teams at Salesforce and Modal already use it to review architecture and specification changes. That’s a claim from the team itself and hasn’t been independently verified.
What limitations does Whiteboard have today?
The README lists them openly:
- You can’t edit files. The founders say they’re evaluating an edit feature, but haven’t yet defined how it should work.
- Reviews spanning multiple repositories aren’t well supported.
- Shared reviews don’t update. If you share a review and then change it, others won’t see the changes until you share it again.
Add to that the beta label, builds only for macOS and Fedora, and a release pace so fast that what you write today can be outdated in weeks.
Is it worth trying Whiteboard?
Try it if you use macOS or Fedora and already let an agent write complete features. Whiteboard doesn’t make the agent write better code. What it does is lower the cost of understanding that code before you merge it, and that’s exactly where most of us are falling behind. If you only need to generate architecture diagrams from scratch, Archify is a better fit.