<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Onecarat Labs/Takeshi Yokoyama</title>
    <description>The latest articles on DEV Community by Onecarat Labs/Takeshi Yokoyama (@onecaratlabs).</description>
    <link>https://dev.to/onecaratlabs</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3960820%2F57056b82-130b-4e04-8ade-76bfd4343967.png</url>
      <title>DEV Community: Onecarat Labs/Takeshi Yokoyama</title>
      <link>https://dev.to/onecaratlabs</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/onecaratlabs"/>
    <language>en</language>
    <item>
      <title>I turned my Claude Code and Codex sessions into a tiny stick-figure office</title>
      <dc:creator>Onecarat Labs/Takeshi Yokoyama</dc:creator>
      <pubDate>Sun, 22 Nov 2026 03:55:02 +0000</pubDate>
      <link>https://dev.to/onecaratlabs/i-turned-my-claude-code-and-codex-sessions-into-a-tiny-stick-figure-office-3goo</link>
      <guid>https://dev.to/onecaratlabs/i-turned-my-claude-code-and-codex-sessions-into-a-tiny-stick-figure-office-3goo</guid>
      <description>&lt;p&gt;Nice to meet you. I'm Onecarat, from Onecarat Labs.&lt;/p&gt;

&lt;p&gt;I have put my introduction on one page. If you are curious, read it as if you were watching a short film: &lt;a href="https://onecarat.dev/my-story/" rel="noopener noreferrer"&gt;ONECARAT — The Story So Far (preview)&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;This time, I made something like this.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuyphkpsqx515dqoqtjem.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuyphkpsqx515dqoqtjem.gif" alt="Onecarat Agent Office, the Claude Code version" width="630" height="393"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu18d8r8r6c119e78g6n0.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu18d8r8r6c119e78g6n0.gif" alt="Onecarat Agent Office CX, the Codex version (display test with synthetic data)" width="620" height="378"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;These are two small windows that show what a coding agent is doing, as stick figures working in an office.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Onecarat Agent Office&lt;/strong&gt; watches Claude Code. The main agent is the manager, and each subagent it starts gets a desk of its own.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Onecarat Agent Office CX&lt;/strong&gt; watches Codex. It is the same idea: a manager, and coworkers who appear when subagents really run.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Both only watch, both run on Windows, and both are free under the MIT license.&lt;/p&gt;

&lt;p&gt;They started from the same idea, but they did not end up the same. The Claude Code office turned out to be about &lt;strong&gt;motion&lt;/strong&gt;: the figures keep moving, the frame blinks when it needs you, and the sky changes with the time of day. The Codex office turned out to be about &lt;strong&gt;stillness&lt;/strong&gt;: a quiet room that shows what it has observed, and says so when it does not know.&lt;/p&gt;

&lt;p&gt;One more thing. Each agent built its own office. Claude Code built the Claude Code version, and Codex built the Codex version. Each one got a first version running on its own, and a fair amount of fixing followed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why stick figures?
&lt;/h2&gt;

&lt;p&gt;I spend most of my days around coding agents, and I think they deserve to be liked a little more than they are.&lt;/p&gt;

&lt;p&gt;For people who are not developers, a terminal is still a scary place. Text scrolls past, something is clearly happening, and it is hard to tell whether the agent is thinking, stuck, or waiting for you.&lt;/p&gt;

&lt;p&gt;I did not want to build another dashboard. I wanted something that makes you smile. A manager sits at a desk, hands work to the team, and everyone starts typing. When a small figure is working hard on your behalf, the session stops looking like a log and starts looking like a team.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you see
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The main agent is the &lt;strong&gt;manager&lt;/strong&gt;. Each subagent it starts gets a desk of its own.&lt;/li&gt;
&lt;li&gt;The figures change with the work: thinking, reading, editing, running commands, testing, researching, delegating.&lt;/li&gt;
&lt;li&gt;One window per session. Run four sessions and you get four small offices side by side.&lt;/li&gt;
&lt;li&gt;The labels are in English or Japanese, following the Windows display language.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The Claude Code version has two extras. When Claude Code is waiting for your approval, a banner appears and the frame blinks orange, so you notice even if the terminal is behind another window.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flf5zpwkydwwetpe9nlaf.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flf5zpwkydwwetpe9nlaf.png" alt="The approval banner and the orange frame" width="630" height="393"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;And the office follows your PC clock: morning, day, evening, night. If your agents are still at their desks at night, you will know.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F50src2nsik0mv9ewn1h8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F50src2nsik0mv9ewn1h8.png" alt="The office at night" width="800" height="203"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What it does not do, on purpose
&lt;/h2&gt;

&lt;p&gt;The window cannot type, send a request, approve, or refuse. It does not save your session, call a model, or talk to any outside service. You keep working in the terminal. The window only watches.&lt;/p&gt;

&lt;p&gt;It did not start that way. Versions up to 0.3.0 ran in a browser tab, let you type requests, and had a "Research Lab" mode where a researcher pinned findings to a board. I removed all of that in 1.0.&lt;/p&gt;

&lt;p&gt;The reason is simple. Claude Code changes quickly, and as one person with a side project, I could not keep a large tool in step with it. A window that only watches depends on very little, so there is less to break. I would rather ship a small thing that keeps working than a big one that is always catching up.&lt;/p&gt;

&lt;p&gt;The old versions are still in the repository's releases if you are curious.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two agents, two ways to watch
&lt;/h2&gt;

&lt;p&gt;This was the most interesting part to build, and it is where the motion and the stillness come from. The two tools look alike, but they learn what the agent is doing in opposite ways.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Claude Code: hooks.&lt;/strong&gt; The Claude Code version is a plugin. Claude Code's hooks tell a small local app what just happened, as it happens. The hooks pass along only what the animation needs. Your prompts, commands, file contents, and tool output stay where they are. The app listens on 127.0.0.1 only and keeps events in memory.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Codex: session logs.&lt;/strong&gt; The Codex version works the other way around. A Node.js process reads the session logs that Codex writes on the same PC, and hands only the type of activity to the window. A coworker appears only when the logs show a real subagent with a real parent.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fk9g49cviag7vxqwjb4s6.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fk9g49cviag7vxqwjb4s6.png" alt="The Codex office with two coworkers at the end of a turn (display test with synthetic data)" width="620" height="378"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Being told and reading afterwards are not the same thing, and the difference shows:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;With hooks, the tool knows when an approval is pending. With logs, it often cannot. In the Codex version I tested (0.169.0), approval requests are not written to the logs, so you still need to check Codex itself.&lt;/li&gt;
&lt;li&gt;Logs can go quiet. If nothing new arrives for 15 seconds, the Codex window says "current state unknown" instead of guessing.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I chose that on purpose. A figure that keeps typing when I do not know what is happening would be a small lie. An office that sometimes says "I don't know" felt better.&lt;/p&gt;

&lt;p&gt;As for letting each agent build its own office: asking an agent to build the window for watching itself turned out to be a good way to get to know it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Claude Code version&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Needs Windows 10 or 11, Claude Code, and Node.js 18 or newer.&lt;/li&gt;
&lt;li&gt;Install it as a plugin with the two PowerShell commands in the README, then restart Claude Code.&lt;/li&gt;
&lt;li&gt;Type &lt;code&gt;Start Onecarat Agent Office&lt;/code&gt; in Claude Code, or run &lt;code&gt;/onecarat-agent-office:office start&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Code: &lt;a href="https://github.com/onecarat-labs/onecarat-agent-office" rel="noopener noreferrer"&gt;https://github.com/onecarat-labs/onecarat-agent-office&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Codex version&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Needs Windows 10 or 11, .NET Framework 4.8, Node.js 24 or newer, and Codex saving its session logs on the same PC.&lt;/li&gt;
&lt;li&gt;For now it is source only. You build and install it yourself, following the setup guide: &lt;a href="https://github.com/onecarat-labs/onecarat-agent-office-cx/blob/main/docs/SETUP.en.md" rel="noopener noreferrer"&gt;https://github.com/onecarat-labs/onecarat-agent-office-cx/blob/main/docs/SETUP.en.md&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;There is no ready-to-run EXE or auto-update yet.&lt;/li&gt;
&lt;li&gt;Code (English README): &lt;a href="https://github.com/onecarat-labs/onecarat-agent-office-cx/blob/main/README.en.md" rel="noopener noreferrer"&gt;https://github.com/onecarat-labs/onecarat-agent-office-cx/blob/main/README.en.md&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Some rough edges, so you are not surprised:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Both are Windows only. There is no macOS or Linux build.&lt;/li&gt;
&lt;li&gt;The Claude Code version is not code-signed, so Windows may show a warning the first time.&lt;/li&gt;
&lt;li&gt;The Codex version has two coworker seats. Extra subagents show up as a count.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Wrapping up
&lt;/h2&gt;

&lt;p&gt;This tool will not make your agent faster or smarter. It only changes how the work feels while you wait.&lt;/p&gt;

&lt;p&gt;But I think that matters. If more people are going to work with agents, the agents need to feel less like a black box and more like someone you can glance at across the room. Stick figures are my small attempt at that.&lt;/p&gt;

&lt;p&gt;If you try it, tell me what your office looked like.&lt;/p&gt;




&lt;p&gt;Code (MIT): &lt;a href="https://github.com/onecarat-labs/onecarat-agent-office" rel="noopener noreferrer"&gt;onecarat-agent-office&lt;/a&gt; · &lt;a href="https://github.com/onecarat-labs/onecarat-agent-office-cx/blob/main/README.en.md" rel="noopener noreferrer"&gt;onecarat-agent-office-cx&lt;/a&gt;&lt;br&gt;
Onecarat Labs: &lt;a href="https://onecarat.dev" rel="noopener noreferrer"&gt;https://onecarat.dev&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claudecode</category>
      <category>codex</category>
      <category>opensource</category>
    </item>
    <item>
      <title>How AI reads your website, and what that means for the people who build it</title>
      <dc:creator>Onecarat Labs/Takeshi Yokoyama</dc:creator>
      <pubDate>Sun, 31 May 2026 09:33:32 +0000</pubDate>
      <link>https://dev.to/onecaratlabs/how-ai-reads-your-website-and-what-that-means-for-the-people-who-build-it-3h64</link>
      <guid>https://dev.to/onecaratlabs/how-ai-reads-your-website-and-what-that-means-for-the-people-who-build-it-3h64</guid>
      <description>&lt;p&gt;&lt;em&gt;By Takeshi Yokoyama — Onecarat Labs&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Hi. I'm Yokoyama, and I build a local-first AI text editor as a side project, along with a few other experimental tools. Working on them, I keep running into the same question about where the web is going. This post is one observation, plus a small experiment I built to test it — including a Chrome extension you can actually try.&lt;/p&gt;

&lt;p&gt;The short version: I think websites will increasingly be read through AI agents, reshaped per reader, on the fly. And once that happens, there's a clear gap between sites that are easy for an AI to read and sites that aren't.&lt;/p&gt;

&lt;h2&gt;
  
  
  What's starting to happen
&lt;/h2&gt;

&lt;p&gt;Until now, people read websites as websites. You open the top page, follow the menu, read the body, click a button — tracing the path the maker designed.&lt;/p&gt;

&lt;p&gt;As local AI and AI agents become normal, that breaks. People stop opening the page directly. They tell an AI what they want — &lt;em&gt;"Can I try this quickly?"&lt;/em&gt;, &lt;em&gt;"I just want to check it's safe"&lt;/em&gt;, &lt;em&gt;"Just the gist"&lt;/em&gt; — and the AI reads the web and reshapes it into the form that reader wants. What the reader receives is no longer the layout the maker built.&lt;/p&gt;

&lt;p&gt;This isn't speculation. The idea that AI generates the interface for the reader already has a name — &lt;strong&gt;Generative UI&lt;/strong&gt; — and it's one of the hottest areas in frontend right now, with Google, Vercel and others building toward it.&lt;/p&gt;

&lt;p&gt;But notice who's holding the pen in almost every version of that story: the &lt;strong&gt;site&lt;/strong&gt;, or an AI &lt;strong&gt;embedded in an app&lt;/strong&gt; — something under the maker's control. What I'm looking at is one step past that: &lt;strong&gt;a local AI, in the reader's own hands, reshaping any site into that person's preferred form — with no involvement from the maker at all.&lt;/strong&gt; The initiative moves from the maker to the reader.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that nags at me as a builder
&lt;/h2&gt;

&lt;p&gt;I build software too. So this shift nags at me.&lt;/p&gt;

&lt;p&gt;A site carries its maker's intent and rights. The order things appear in, what gets emphasized, the tone. Design, copy, flow — all of it is deliberate. Having an AI quietly reorder, rewrite, and re-tone that doesn't sit right with me.&lt;/p&gt;

&lt;p&gt;But there's no escaping the other side of it: &lt;strong&gt;a site that AI can't read well might as well not exist.&lt;/strong&gt; Once people search and decide through AI, a site the AI couldn't parse drops out of the running before the reader ever sees it. However carefully you wrote the HTML, if it's too noisy for the AI to read, that care turns into lost opportunity.&lt;/p&gt;

&lt;p&gt;So there's a bind. Being reshaped without consent isn't what I want. But not being read means losing the chance entirely. Both are true at once.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where this conversation already is
&lt;/h2&gt;

&lt;p&gt;I should be honest: I'm not the first to circle this. "Sites should carry machine-readable information for AI agents" is an active conversation abroad, under the name &lt;strong&gt;agentic web&lt;/strong&gt;. Two reference points:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;llms.txt&lt;/strong&gt; (proposed by Jeremy Howard in 2024) — a Markdown file at your site root pointing AI at your important pages. As of 2026, adoption sits around 10%, and the major crawlers barely fetch it — yet it's increasingly framed as a "Business-to-Agent" play, and IDE agents like Cursor and Claude Code do read it routinely. Interestingly, several analyses expect the ecosystem to move &lt;em&gt;from one site-level file toward per-page, machine-readable signals.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;NLWeb&lt;/strong&gt; (Microsoft, announced at Build 2025) — turns a site into something an AI can query in natural language, reusing the Schema.org structured data a site already has. Every NLWeb instance also acts as an MCP server.&lt;/p&gt;

&lt;p&gt;So the broad direction — &lt;em&gt;a human web and an AI-facing surface, side by side&lt;/em&gt; — isn't a hypothesis anymore. It's already moving.&lt;/p&gt;

&lt;h2&gt;
  
  
  So what do I think is actually new?
&lt;/h2&gt;

&lt;p&gt;Given all that, here are two gaps I don't think the existing efforts fill — and they're the point of this post.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;First, the existing work is mostly about the "make me discoverable" side.&lt;/strong&gt; llms.txt and NLWeb are largely about getting a &lt;em&gt;cloud-side&lt;/em&gt; AI search or agent to find and cite your site — the owner pushing information out. What I'm looking at is the opposite end: &lt;strong&gt;a local AI in the reader's hands rebuilding the page's display for that one person's intent.&lt;/strong&gt; The reader/viewer side. Most of the conversation leans toward "help agents discover the site"; the "the reader's local AI re-renders it per intent" side is still thin.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Second, there's almost no notion of how AI-driven reshaping coexists with the original's authority.&lt;/strong&gt; In an era where AI summarizes, reorders, and re-tones a site, how do you preserve the maker's intent? That question is barely present in llms.txt or Schema.org. Everyone talks about how the &lt;em&gt;view&lt;/em&gt; gets built; almost no one pairs it with &lt;em&gt;"and the original stays sovereign."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;My answer to the bind is to split the site into two layers — and hold them both at once.&lt;/p&gt;

&lt;h2&gt;
  
  
  My answer: two layers, held together
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;The original (your HTML) is untouchable.&lt;/strong&gt; The maker controls it completely. Rights and intent live here. If you want it as the maker intended, you read the original.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The derived view is the AI's to build freely.&lt;/strong&gt; Reorder, summarize, change tone — fine. Because it doesn't touch the original. The original still stands, as the maker made it. The derived view is just one reader's way of looking, for themselves. Don't like it? Read the original.&lt;/p&gt;

&lt;p&gt;Here's the thing that makes the free part okay: &lt;strong&gt;because the original is preserved, the derived view is free to be whatever it needs to be.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  So what does the maker actually do?
&lt;/h2&gt;

&lt;p&gt;This is the practical bit. If the derived view is the AI's to build, the maker doesn't need to design that view. The maker does exactly one thing: &lt;strong&gt;make the site easier for AI to read.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Force an AI to parse human-facing HTML and you get misreadings, and instability. Better to hand over the site's information, up front, in a form the AI can read cleanly. So I describe each page as &lt;strong&gt;material&lt;/strong&gt; — not layout — in a JSON file that sits beside the HTML. One extra line in the &lt;code&gt;&amp;lt;head&amp;gt;&lt;/code&gt; points to it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight html"&gt;&lt;code&gt;&lt;span class="nt"&gt;&amp;lt;link&lt;/span&gt; &lt;span class="na"&gt;rel=&lt;/span&gt;&lt;span class="s"&gt;"ai-source"&lt;/span&gt; &lt;span class="na"&gt;type=&lt;/span&gt;&lt;span class="s"&gt;"application/json"&lt;/span&gt; &lt;span class="na"&gt;href=&lt;/span&gt;&lt;span class="s"&gt;"index.ai.json"&lt;/span&gt; &lt;span class="nt"&gt;/&amp;gt;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json-doc"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"0.1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"site"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt;    &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Onecarat Labs"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://onecarat.dev"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"purpose"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"..."&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"page"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt;    &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"path"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"index.html"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"title"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"OnecaratEditor"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"product"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"summary"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"..."&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"sections"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"download"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"title"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Download"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"intent"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"try-now"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"content"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Plain-text summary..."&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"actions"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt;  &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"download-latest"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"label"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Download"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"..."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"intent"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"try-now"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The point: this JSON is &lt;strong&gt;not&lt;/strong&gt; an instruction telling the AI how to display anything. It's material, handed over so the AI can read cleanly. The freedom to summarize, reorder, re-tone stays with the reader's AI. The maker doesn't constrain it.&lt;/p&gt;

&lt;p&gt;And the result: doing this &lt;strong&gt;makes the site easier for AI to read.&lt;/strong&gt; Easier to read means less lost opportunity. The original HTML stays untouched; whoever wants it reads it. The maker only builds the on-ramp. How the view gets assembled is left to the AI.&lt;/p&gt;

&lt;p&gt;It's closer to SEO than to a rendering engine. A site owner never tries to build Google's results page &lt;em&gt;for&lt;/em&gt; it. They just want to be found, so they make the site easy to crawl. How the results page is assembled is Google's job. This is that — with an AI agent where the search engine used to be. (And note: the ecosystem is reportedly drifting from one site-level file toward &lt;em&gt;per-page&lt;/em&gt; signals anyway — which is exactly the one-JSON-per-page shape here.)&lt;/p&gt;

&lt;h2&gt;
  
  
  Onecarat Lens — feeling "it got read"
&lt;/h2&gt;

&lt;p&gt;Words only go so far. So I built something you can run: &lt;strong&gt;Onecarat Lens&lt;/strong&gt;, a Chrome extension. It's open source.&lt;/p&gt;

&lt;p&gt;🔗 &lt;strong&gt;&lt;a href="https://github.com/onecarat-labs/onecarat-lens" rel="noopener noreferrer"&gt;https://github.com/onecarat-labs/onecarat-lens&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Lens reads the &lt;code&gt;ai-source&lt;/code&gt; a site provides, and lets a local AI — Chrome's built-in Gemini Nano — assemble a view for the reader's intent, painted as an overlay &lt;strong&gt;on top of&lt;/strong&gt; the original. Concretely it:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;finds the &lt;code&gt;ai-source&lt;/code&gt; link on the page (does nothing if there isn't one),&lt;/li&gt;
&lt;li&gt;asks the reader for intent — preset buttons (&lt;em&gt;try-now / safety / overview / details / choose&lt;/em&gt;) plus a free-text box,&lt;/li&gt;
&lt;li&gt;lets the local AI decide what to keep, how to order it, how to reword it,&lt;/li&gt;
&lt;li&gt;paints the result as an overlay that never replaces the original,&lt;/li&gt;
&lt;li&gt;and gives a toggle to flip back to the untouched original anytime.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fewj6e76bc1svl8yjlevb.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fewj6e76bc1svl8yjlevb.png" alt=" " width="800" height="407"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;On the same page, &lt;em&gt;"I just want to try it"&lt;/em&gt; pushes the download and quick-start to the top; &lt;em&gt;"I want to check it's safe"&lt;/em&gt; pushes the local-first, no-cloud, code-signing notes up instead. Same site, different shape — and the source is material the site offered, not a guess scraped from HTML.&lt;/p&gt;

&lt;p&gt;What I want you to feel here isn't the polish of the view. It's the fact that &lt;strong&gt;the site got read&lt;/strong&gt; — cleanly, on the reader's terms — and that the original is still right there, one toggle away.&lt;/p&gt;

&lt;p&gt;I'll be honest about the rough edges. Gemini Nano is small, so when it's pushed it produces clumsy summaries, or falls back. Lens shows that fallback openly rather than hiding it. That's deliberate. The point of this experiment is honesty about what local AI can and can't do yet.&lt;/p&gt;

&lt;h2&gt;
  
  
  A note on where this came from
&lt;/h2&gt;

&lt;p&gt;This whole idea came out of a wrong turn. An earlier version of Lens read a site's raw HTML and let the AI rebuild it directly. And it felt like overriding the maker's intent. The fix wasn't to make the AI more polite about it. It was to let the original stay sovereign, and have the site offer material for the AI on its own terms. That realization is what &lt;code&gt;ai-source&lt;/code&gt; is.&lt;/p&gt;

&lt;h2&gt;
  
  
  Wrapping up
&lt;/h2&gt;

&lt;p&gt;"HTML for humans, JSON for AI" is no longer just my hypothesis — it's already moving, as llms.txt and NLWeb. What I think is new is two things: not the discovery side but the &lt;strong&gt;reader's local AI re-rendering per intent&lt;/strong&gt;, and a maker who declares nothing about the view but simply &lt;strong&gt;makes the site readable&lt;/strong&gt;, paired with an original that stays sovereign.&lt;/p&gt;

&lt;p&gt;Local AI will only get more practical. But it holds no fresh information, so it'll be asked to go and read the web. And today's human-facing web is too noisy for it. So the site offers something readable; the original stays whole; the reader sees it through their own local AI, in the shape that fits them.&lt;/p&gt;

&lt;p&gt;Onecarat Lens is a small experiment to test that future. It doesn't replace the web. It adds an on-ramp for AI to the human-facing HTML — opened from the maker's side, deliberately, on the maker's terms.&lt;/p&gt;

&lt;p&gt;If local AI agents become normal, sites will be read by AI too. Whether you've made yours readable, by then, is the thing that will matter.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Code: &lt;a href="https://github.com/onecarat-labs/onecarat-lens" rel="noopener noreferrer"&gt;github.com/onecarat-labs/onecarat-lens&lt;/a&gt; · Built by &lt;a href="https://onecarat.dev" rel="noopener noreferrer"&gt;Onecarat Labs&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>localai</category>
      <category>webdev</category>
      <category>chrome</category>
    </item>
  </channel>
</rss>
