Whiteboard: open-source IDE where humans and coding agents design on a shared canvas

Whiteboard, a desktop IDE from a team in Y Combinator's W26 batch, arrived as an open-source project via a Show HN post on Hacker News on September 24. Its premise is that as coding agents take over more of the writing, developers need a workspace built for reviewing and steering that work rather than simply editing files, so Whiteboard places the human and the agent on the same canvas.

Humans and agents on one canvas

According to the project's README, Whiteboard plugs into agent tools developers already use, such as Claude Code and Codex, rather than shipping its own model integrations. Once connected, the agent receives an SDK that lets it draw on an in-app canvas to explain its work, whether a proposed API, a diagram or a summary of a change. The suggested starting point is asking an agent to review your current branch against an up-to-date main branch and open the result in the app. Anything on the canvas you dislike can be highlighted and handed back to the agent to redraw.

The team's guidance notes the tool currently works best with frontier models, naming GPT-6 Sol and Claude Opus 5.5 for their balance of capability, cost and speed. The developers add that they are still baking sensible defaults into the system prompt, which they cite as one reason for open-sourcing the project.

Diagrams that connect back to code

The developers argue that pure diagramming tools make it hard to tie a spec to real code, especially since tradeoffs often only become visible after a first implementation pass. In Whiteboard, clicking a visualization, such as a sequence diagram, an entity-relationship diagram or a quote lifted from the agent's trace, jumps straight to the underlying source. Because the app builds on VS Code, code navigation ships with familiar keybindings and LSP support out of the box.

An AST-aware diff viewer

Raw diffs get noisy quickly, so the team wrote a semantic diff viewer in Rust that is aware of the abstract syntax tree. Large newly added functions are summarized as pseudocode, and unit-test and documentation churn is collapsed by default, so only relevant changes surface. The behavior can be customized through a WASM-based plugin system.

A log of what the agent decided

To make autonomous work auditable, Whiteboard gives agents tools to query and link their own traces on the canvas. The goal, per the README, is to visualize the requirements you set, see how they were implemented and identify which decisions the agent made on its own, reasoning that normally stays buried in logs.

A vendored Code OSS fork

Whiteboard is MIT-licensed and operates entirely on your local repositories. The team explains that it vendors Code OSS outright rather than maintaining a patch set, because patch files are awkward for coding agents to work with and stock VS Code carries a lot the project does not need, with the developers estimating that roughly 45 percent of the codebase is Copilot-related. Upstream security and feature patches are merged in as they appear. A hosted product for teams is planned, but the developers say self-hosting will always remain an option.

The team's reasoning for starting from VS Code at all: with agents doing most of the typing inside TUIs and desktop apps, people now mostly reach for text editors to review diffs, so an editor designed for review seemed overdue.

Known gaps and telemetry

The README is frank about limitations. You cannot edit files inside Whiteboard, reviews spanning multiple repositories are poorly supported, and once a review is shared, later updates do not reach the recipient, so it has to be shared again. On privacy, the team says anonymous telemetry excludes code, diffs, canvas content, prompts and model output, and can be turned off at any time.

Why it matters

Agent tooling has largely optimized for generating code faster. Whiteboard bets that the real bottleneck is comprehension, a view it grounds in essays it cites by Geoffrey Litt on understanding as the constraint in modern software engineering and by Maggie Appleton on AI-generated code becoming tomorrow's legacy code, alongside Karpathy's writing on LLM agents. If agents write most software, the review surface becomes the primary interface between human intent and machine output, and a shared canvas with traceable decisions and code-linked diagrams is a credible shape for that interface. The MIT licence and local-first posture also lower the barrier for teams unwilling to send their code to a hosted review tool.