The code arrived clean, functional, passing all tests. But understanding what automated coding assistants actually change and why has become a common challenge for development teams.
This problem—understanding the handiwork of automated coding assistants—has become common enough that a four-person startup thinks it can build a business around solving it. /dev/fast, a company in Y Combinator's Winter 2026 batch, released Whiteboard on September 24, 2026 as an open-source desktop application designed to translate the opaque output of AI coding agents into something human developers can actually review.
The tool sits between the agent and the developer, converting raw code changes into architecture diagrams, decision logs, and what the team calls "semantic diffs" that surface intent rather than just altered lines. /dev/fast says that Salesforce and Modal, a serverless platform company, have both begun testing Whiteboard internally.
Whether the tool gains traction beyond early adopters remains an open question. The GitHub repository collected 490 stars within a day of its Show HN post on Hacker News, which as of September 25, 2026 had drawn 272 upvotes and 110 comments. But stars and upvotes measure curiosity more than commitment.
The Review Problem
As AI coding agents like Claude Code and Cursor write larger portions of production code, developers face a peculiar challenge. The code often works. It may even be elegant. But the reasoning behind architectural decisions gets lost in translation, buried in verbose agent traces or scattered across dozens of modified files.
Whiteboard runs atop Code OSS, the MIT-licensed upstream version of Visual Studio Code, and displays agent traces alongside the code they modify. When an agent makes a change, the tool generates sequence diagrams, entity-relationship diagrams, and trace quotes linked directly to the underlying code through language server protocol support.
The company's core technical bet is diffr, a Rust-based abstract syntax tree diff viewer built as a fork of difftastic. It collapses unit tests and large documentation blocks, then summarizes new functions as pseudocode rather than showing every line. "Agents struggle with patch sets," the team wrote in the repository's README, explaining their decision to build a custom diff engine with a WebAssembly plugin system instead of relying on existing tools.
This is perhaps more technical overhead than some teams will tolerate. The tool requires local installation and doesn't yet support multi-repository browsing or live-synced review updates, according to limitations documented in the GitHub repository.
Team Background
CEO Sidharth Menon previously worked on cloud infrastructure at Palantir and studied computer science at Harvard. CTO Milan Bhandari co-founded Bolto, a YC Summer 2023 company, and also came from Palantir. Ketan Agrawal built data-querying agents at Snowflake before working on AI red-teaming at Robust Intelligence and attending Stanford. Alex Iansiti, the chief product officer, worked at Flowcode and Pinterest.
The Palantir connection runs deep—two of the four founders spent time there, which may explain the tool's focus on making opaque systems legible. Palantir built a business on exactly that premise, though in intelligence analysis rather than code review.
Where It Fits
Whiteboard competes in a crowded space where developers already use Cursor for AI-assisted editing, Claude Code for terminal-based coding, and GitHub Copilot CLI for command-line assistance. The company positions its tool as a review layer rather than an orchestrator, designed to work alongside those agents.
Menon outlined a workflow in the Show HN thread where teams might use Greptile to automate review of trivial changes and escalate larger architectural decisions to Whiteboard for human judgment. One developer filed a GitHub issue requesting Copilot CLI integration. Menon replied that the team was planning to add the feature, though the integration had not yet been confirmed at the time.
The tool ships as a direct download for macOS on Apple silicon and as version 0.1.2 for recent Fedora Linux distributions, with installation instructions documented on the company's website. It carries an MIT license and runs against local repositories. The company has indicated that a hosted product for teams is under consideration, but stated in documentation that "everything will always remain self-hostable."
The Editing Question
Whiteboard does not currently support file editing, which sets it apart from most developer tools in this category. Developers can review what agents did but can't immediately fix problems they spot. Menon wrote in the Show HN thread that the team is "strongly considering adding an editing feature" while maintaining focus on architecture-level review.
That constraint may prove either clarifying or limiting. Some teams might appreciate a tool purpose-built for review. Others may find it frustrating to switch contexts when they spot an issue that requires a quick fix. The market will likely decide which camp is larger.
