We benchmarked an 84% token reduction. Then we open sourced the protocol. — Stacklist Blog

Stacklist Team

May 11, 2026· Company Updates

We benchmarked an 84% token reduction. Then we open sourced the protocol.

Meet ACP (Atomic Content Protocol). The open standard designed to help AI agents read content with far fewer tokens.

Why We Built ACP and Why We're Open Sourcing It - YouTube

Tap to unmute

Why We Built ACP and Why We're Open Sourcing It Stacklist

Stacklist57 subscribers

Watch on

The internet was built for humans. But the fastest-growing users of the internet are now AI agents. They read the same pages over and over. They burn tokens trying to figure out what each piece of content is. There's no shared language for any of it.

That's the problem we kept running into while building Stacklist. First for our users, then for our own internal AI tools. And at some point it became clear that the fix wasn't another feature, or a smarter parser, or a proprietary integration. It was a standard.

So we built one. It's called ACP, the Atomic Content Protocol, and we're open sourcing it today.

84% fewer tokens, live benchmark

The proof of why ACP matters is the benchmark Martina ran during the Building Stacklist episode above.

She enriched a stack of nine cards with ACP, then had an agent read them, first cold, then again with the enrichment in place. The second read consumed roughly 200 tokens per card instead of 20,000. A consistent 84% reduction across the set, just from the agent knowing what the content was before reading it.

At single-card scale, that's an optimization. At network scale, it's the difference between agentic search and discovery being affordable and being a luxury.

A protocol, not another app

The reason a number like 84% is even possible comes down to what ACP actually is. The clearest way to explain it is the same way Kyle explains it in the episode: SMTP.

Before SMTP, email providers were walled gardens. AOL couldn't talk to Yahoo. Hotmail couldn't talk to Gmail. Every system had its own way of structuring messages, and the only thing that made the modern email network possible was a shared protocol. A simple, open agreement on how to pass mail back and forth.

Content on the internet has the same problem right now, just one layer up. Notion, Obsidian, Slack, a saved markdown file, a recipe site, a real estate listing… Each one structures its content differently, and each one forces an agent to spend tokens figuring out what it's looking at before it can do anything with it.

ACP is the shared agreement. It's the layer of enrichment that tells an agent: this is what this content is, this is how it relates to other content, this is the confidence we have in it. JSON-LD already does some of this for crawlers. ACP does it for agents.

Stacklist is the first app on ACP

Every card and stack on Stacklist is already structured the way ACP specifies. When you save something via our MCP, the browser extension, or the app, the card automatically picks up the enrichment layer: structural relationships, agent-readability flags, token metadata, and confidence scoring. We're not bolting this on later. Stacklist was built to be "LLM delicious" from day one.

That's also why a Stacklist hub works the way it does for AI discovery. The content isn't just findable by humans browsing. It's findable by the agents increasingly doing the browsing on behalf of humans.

Open source from day one

For ACP to do what protocols are supposed to do, let different systems talk to each other without negotiating a language every time, it has to be open. Anyone can build with it. Anyone can extend it. Anyone can disagree with how we've structured something and propose a better way.

The success metric isn't Stacklist adoption. It's whether other developers and platforms find ACP useful enough to build on. If that happens, the protocol does what Wikipedia does: it disappears into the infrastructure, and people just use it.

How to start

Two paths in, depending on how deep you want to go. You can find everything you need in this stack.

• Watch the full conversation in the episode embedded above. Kyle and Martina walk through the why, the SMTP analogy, the library/treasure-map framing, the live benchmark, and a tour of an ACP-enriched card and stack.
• Try it today at atomiccontentprotocol.org. The npm package is live, the MCP is shipped, the playground works with your own Anthropic API key, and the setup guide will get you running in a few minutes.

If you build something on ACP, tell us. If you think we got something wrong, also tell us. We’d love to hear from other builders.

More articles

When should a startup startcaring about AI search?Before the website\ \ October 1, 2026 · AI Discovery When should a startup start caring about AI search? Before the website A new company starts with zero online presence. Here's why publishing a Day-One Library: tools, partners, and honest comparisons; beats building a website first. Is AI-search optimizationjust new SEO? What carriesover, what breaks\ \ September 29, 2026 · AI Discovery Is AI-search optimization just new SEO? What carries over, what breaks AI search didn't replace SEO, it changed the scoring unit. Here's exactly which SEO practices still matter, which break, and which just change shape. Find the questions AI can'tanswer yet (and own themfirst)\ \ September 24, 2026 · AI Discovery Find the questions AI can't answer yet (and own them first) Every obvious topic in your niche is already covered. Here's how to find the specific questions AI still answers badly in your market and own them first.

Ask Stacky