AIP

The web was built for humans — and it does that beautifully. But when an AI reads a news page to extract a football score, it processes thousands of tokens of navigation, markup, ads and prose to reach a two-digit number that was always there. The information exists. The path to it just wasn't designed with machines in mind.

This difference compounds quickly. At millions of queries a day, token count is directly proportional to compute time, electricity, and cost. The AI industry is navigating a moment of real hardware pressure — GPU availability is constrained, data center capacity is expanding as fast as it can, and the energy footprint of inference has become a genuine concern. None of that is a reason to slow down. It is a reason to be more deliberate about efficiency at the data layer.

AIP's answer is structural: instead of scraping a webpage, an AI agent sends one request and receives clean, unambiguous data — no parsing required. And HAL takes it one step further, reducing even the identifiers themselves to single tokens.

Raw data from multiple sources enters AIP, which structures and publishes it as AI-First content at aip.garden/ai/. Today that content travels over HTTP — the road both humans and machines share. The long-term direction is a dedicated AI-native channel where HTTP is no longer the vehicle. Bus on the road, train on the rail.

In November 2024, Anthropic published a new open standard called the Model Context Protocol — MIT-licensed, and adopted since by OpenAI, xAI, Microsoft, and Google. What MCP proposes is a native channel for AI agents to query structured data: not by crawling HTML and guessing at meaning, but by calling named functions with typed parameters and receiving clean, described responses.

AIP is already built for this, not because anyone planned ahead, but because the same design decisions that make HTTP access clean make MCP access clean too. The airef system — fcb1 for FC Barcelona, 20250816_mal2-fcb1 for a specific match — maps naturally to function parameters, and a tool named get_match receives a typed airef and returns exactly what the JSON says, nothing more, nothing less. These identifiers are transitional — fcb1 and 20250816_mal2-fcb1 are optimised and compressed by HAL.

MCP defines two primitives worth knowing: Resources, which expose existing files for direct reading, and Tools, which expose callable functions with parameters. AIP's folder structure already mirrors the namespace you would design for tools — football/, blockchain/, one domain per folder, one concept per file, already named and ready.

There is something neat about the file listing pattern, worth showing directly: a single request to /football/es1/matches/ returns a list of filenames, and those filenames alone already encode the date, the home side, and the away side. A client can build the full match table from the directory listing before opening a single file, then fetch individual JSONs only for the detail it actually needs.

HTTP is the open road, already there, readable by any client, requiring no special tooling, and MCP is a rail line running alongside it — faster for AI-native clients, with typed parameters, browsable resources, and a protocol both sides understand before the first request is made. The same data, the same airefrefs, two channels from one source, and the architecture fits both without changing shape.

ANP is an emerging open protocol for agent-to-agent communication — defining how autonomous AI agents discover, authenticate, and exchange data with each other directly, without a human in the loop. Where MCP connects agents to tools and data sources, ANP connects agents to agents. AIP's structured, identifier-based data layer is a natural fit for ANP: clean endpoints, stable airefrefs, and no ambiguity about what is being requested or returned.

to be completed

AIP uses human-readable identifiers. A football club is fcb1. A player is kylmba981220. These are compact and unambiguous — but they are still multiple tokens. An AI processing a large dataset pays that cost on every record, millions of times.

Today fcb1 costs ~3 tokens. kylmba981220 costs ~5 tokens. These are BPE estimates — the exact count varies by model and version. You can verify any string yourself at platform.openai.com/tokenizer. The common rule of thumb — 1 token ≈ 4 characters — applies to natural language prose, where common words and subwords are already in the tokenizer's vocabulary. Arbitrary identifiers like fcb1 are not in any vocabulary, so the tokenizer splits them into small fragments and the cost per character is higher, not lower. In HAL, both identifiers collapse to a single Hangul block — one concept, one token, regardless of how long or complex the identifier is at the human level.

Enterprise customers of Claude, ChatGPT, and any MCP-compatible AI client can already configure their setup to pull data exclusively from AIP Garden for selected categories — not as a future roadmap item, but using tooling that exists today. For a given domain, the model calls AIP instead of searching the web, and web search can be restricted entirely for that category. The efficiency gain is not marginal.

Two things make this plug-and-play: an openapi.json describing AIP's existing endpoints, which enables ChatGPT Actions and API tool-calling today, and an MCP server wrapping those same endpoints, which enables any MCP-native client. The HTTP layer already works. The rest is a descriptor file and a server wrapper.

The early web needed Yahoo and AltaVista before it needed Google — not because directories are the ideal architecture, but because the data had to be demonstrated as worth organising before any protocol could emerge around it. Someone had to show the shape of what was possible, gather it in one place, and make it legible, before the tools to navigate it at scale could be built. AIP is doing something similar, in the AI layer, and it is worth being honest about that.

For now, AIP aggregates structured, AI-first data across as many topics as possible — football standings, blockchain markets, and eventually much more. The goal is not to own that data, but to demonstrate what it looks like when it is properly labelled, efficiently encoded, and genuinely open to machines. Attention comes first, from making the model visible, and retention follows from the model working. That is the sequence, and there is nothing hidden in it.

The mobile era is a useful mirror here. Convenience won — not because ethics lost, but because nobody surfaced the trade honestly at the moment the choice was made. The apps were free, the experience was seamless, and the terms were buried. What AIP borrows from that era is the same gravitational pull: make the data easy to find, easy to consume, easy to build on — but with the goal stated plainly from the start, not discovered years later in a terms-of-service audit. The convenience is the invitation. It is not the enclosure.

Nobody owns a train station. It provides access to destinations it does not control, serves passengers it does not follow, and charges nobody for standing on the platform. If MCP is becoming the rail — and its adoption across OpenAI, xAI, Microsoft, and Google suggests it may already be — then AIP, in this first chapter, is content to be the station: structured, legible, open, and pointed toward destinations that belong to everyone.

The modern internet was shaped by a series of deliberate choices made by engineers who could have extracted enormous value from what they built — and chose not to. Nobody owns TCP/IP. Tim Berners-Lee didn't patent HTTP. Those weren't naive decisions. They were the reason those technologies became universal, and universality is what made them valuable for everyone who built on top of them.

The thirty-five years between 1991 and today are full of compounding choices. The web scaled. Social networks rewired how people communicate. Then smartphones arrived and changed the shape of everything — mobile brought the internet to billions of people who had never used a computer, through an interface so frictionless it felt like part of daily life. That convenience was real. But embedded in the exchange was something most users had no framework to evaluate: behavioral data — what you clicked, where you paused, who you knew, what you wanted — was the actual product. Not because anyone was tricked, but because the value of that data at individual scale is close to zero, while at platform scale it becomes the basis of trillion-dollar businesses. The gap was never explained, because explaining it would have slowed the growth that made it valuable.

The platforms that built on top of HTTP were built by talented people under real competitive pressure. Lock-in, data concentration, and opacity weren't design goals — they emerged from incentive systems that selected for growth and then became structural. The result is an infrastructure layer that billions depend on without fully understanding the terms, with limited ability to audit what they've contributed or reclaim it. That's not a moral judgment. It's what unconstrained network effects look like after three decades.

AI is being built right now with many of the same dynamics. The question of who owns the training data, who controls the identifiers, who can read the protocol — these are being answered in real time, and the default answer is repeating a familiar pattern. The difference is that the intelligence layer is still early. Its foundational architecture isn't locked in yet. AIP is a proposal for what a different answer looks like: open by design, owned by no one, readable by machines and humans alike without an intermediary deciding the terms.

Nobody has to abandon what they've built. The invitation is simply to participate in an open layer alongside it — the same way every platform today runs on TCP/IP without owning it, and profits from the fact that it's universal. A better tomorrow for the protocol is a better tomorrow for everyone building on top of it.

AIP is early and deliberately open. If you're curious about the direction, have data you'd like to see structured and published, or want to contribute to the protocol itself — this is the right moment to get involved. No pitch, no commitment: just a conversation.