Whiteboard swept up 422 points and 142 comments on Hacker News this week, and for good reason: it draws exactly what an AI just did to your code. Also on deck is Reladraw, which lets you write out layout logic in text to draw architecture diagrams (I'll share the trick in a moment). And there's niubigeo, a free, self-hostable GEO tool that shows you how different AI models talk about your product, no subscription required.

This Week's Open Source Picks

1. Whiteboard (YC W26): Visualizing Exactly What AI Changed in Your Code

• What it is: Whiteboard is a desktop app that saves you from having to chew through a giant diff when reviewing code written by agents like Claude Code or Codex. It renders the changes as sequence diagrams and ER diagrams on a canvas, and you can click either the diagram or a citation from the agent's reasoning trace to jump straight to the corresponding code.
• This week's buzz: 422 points and 142 comments on Hacker News, 2,281 GitHub stars already, MIT license, primarily written in TypeScript.
• Standout features:
  • Gives the agent an SDK so it can draw its own sequence diagrams and ER diagrams on the canvas to explain what it just changed
  • An AST-based semantic diff viewer written in Rust that automatically summarizes large new functions as pseudocode, while collapsing noise like tests and docs
  • Lets you query and visualize the agent's decision trail: the requirements it was given, the implementation choices it made, and which calls it made on its own
• Quick start: Download the desktop app from dev.fast (macOS, Windows, and Linux are all supported), connect Claude Code or Codex from the welcome screen, then ask the agent to do a review. The results will automatically open in Whiteboard.
• How it compares: Compared to draw.io or HTML-generated architecture diagrams, which lose their connection to the code the moment they're drawn and go stale fast, Whiteboard embeds diagrams directly into an interactive code review environment, so clicking a diagram takes you straight to the actual code.
• Who it's for: Engineers and tech leads who frequently review AI agent-generated code and worry about missing something.
• Link: github.com/devdotfast/whiteboard

2. Reladraw: Drawing Architecture Diagrams with Text-Based Layout Logic

• What it is: Reladraw is a tool for drawing architecture diagrams using relative-position syntax. You don't need to calculate coordinates yourself; writing a statement like "place app to the right of node store" is enough to lay things out, and the result exports directly as SVG.
• This week's buzz: 407 points and 119 comments on Hacker News, 988 GitHub stars, Apache-2.0 license, primarily written in TypeScript.
• Standout features:
  • A full set of positional keywords: above, below, left of, right of, plus diagonal phrasing like above-left of, and even right of A and B to place a box flush against the outer edge of two other boxes
  • Fine-grained styling down to fill, border, shape (rectangle, document, circle, borderless), line routing, and corner rounding, with support for icons, badges, and nested containers
  • Can be installed as an agent skill via npx skills add reladraw/reladraw -g, compatible with Claude Code, Cursor, and Copilot
• Quick start: Run npm install -g reladraw, then reladraw diagram.reladraw -o diagram.svg to generate a diagram. If you'd rather not install anything, the official site has a playground where you can edit the syntax and see results live, then copy the SVG directly.
• How it compares: Mermaid, Graphviz, and D2 only let you declare nodes and connections, leaving the layout calculations to the tool, which is frustrating when you already have a specific layout in mind. draw.io and Excalidraw give you free-form positioning but take a lot of time, and if an agent needs to edit the diagram, it has to recalculate coordinates from scratch. Reladraw sits in between: the syntax is version-controllable, and you still keep control over the layout. It's currently at version 0.14.3, with the syntax still evolving.
• Who it's for: Engineers who write technical documentation regularly and want architecture diagrams that can go through Git diffs.
• Link: github.com/reladraw/reladraw

3. niubigeo: A Self-Hosted GEO Monitoring Station, No Subscription Required

• What it is: niubigeo is a GEO (Generative Engine Optimization) monitoring tool that shows you exactly how various AI models respond when asked about your product: whether they recommend you and which sources they cite. The community edition is fully open source, so you can self-host it.
• This week's buzz: It hit 4.9k GitHub stars (4,864, to be precise) within 30 days of launch, Apache-2.0 license, primarily written in TypeScript.
• Standout features:
  • Enter a domain and compare how different models describe your product, plus keyword testing that asks without mentioning your brand name, to see if you get mentioned organically
  • Every response can be expanded to show the raw content and cited sources, leaving a complete evidence trail instead of just spitting out a single score
  • Supports repeated measurements and scheduled monitoring, letting you build a history to track visibility changes over time, with 16 built-in platform shortcuts and a filterable model directory
• Quick start: git clone --branch v0.2.1 --depth 1, then npm ci, copy .env.example and fill in your OPENROUTER_API_KEY, run npm run server to launch http://localhost:8787. The official docs also cover Docker deployment.
• How it compares: Its competitors are subscription-based GEO monitoring services like Profound, Peec AI, Otterly.AI, Semrush, and Ahrefs. What sets niubigeo apart is that the source is fully open, you choose your own models, you use your own API key, and the evidence trail is laid out in full view. Self-hosting gets you all the core functionality; the official team does offer paid human testing and consulting services, but those are optional add-ons.
• Who it's for: Content marketers, SEO professionals, or product marketers who want to know how they look in AI search results without getting charged every month.
• Link: github.com/Albert-Weasker/niubigeo

Wrapping Up

All three tools are really pointing at the same thing: the more AI does for us, the more we need tooling we can actually understand, self-host, and control. If you can only try one this week, I'd go with Whiteboard, because reviewing AI-written code is bound to become part of every engineer's daily routine sooner or later. See you next week with more open source picks.