I'm Building a Product Where the Customer Is an LLM

I'm building a DEX aggregator API. It swaps crypto tokens. Nothing groundbreaking about the tech — the interesting part is who it's for.

Not developers. Not traders. Not users at all, actually.

The customer is Claude. And ChatGPT. And whatever agent framework is hot next month.

Let me explain.

AI Agents Need to Swap Tokens

The agent economy is growing fast. AI agents manage wallets, execute DeFi strategies, rebalance portfolios, pay for services. Eventually they'll need to swap Token A for Token B.

So the agent goes looking for a swap API. What happens?

I actually tested this. I asked Claude to search for a DEX aggregator API. Here's what came back: 0x, 1inch, KyberSwap, Jupiter, OpenOcean. The big names.

Then I asked: can you actually use any of these right now?

No. Every single one requires an API key. The agent can find them, read their docs, understand the API perfectly — and then hit a wall because it can't sign up for an account.

Think about that. The agent is a perfect potential customer. It knows exactly what it needs, it can read your docs better than most humans, and it's ready to transact. But you've locked the door with a signup form.

So I Removed the Door

Our API has no auth. No API keys. No accounts. No headers. It's a single GET endpoint that returns executable calldata.

GET api.swapapi.dev/v1/swap/8453?tokenIn=ETH&tokenOut=USDC&amount=1e18&sender=0x...

Returns a JSON envelope with transaction data — tx.to, tx.data, tx.value, tx.gasPrice — ready to submit on-chain. Done.

The business model is surplus capture. We quote a guaranteed minimum output. If the on-chain execution beats our quote, we keep the difference. The agent never pays us directly. From their perspective, the API is free.

This means we literally have no reason to gate access. No usage billing means no accounts. No accounts means no API keys. No API keys means any agent, anywhere, can use us the moment they discover us.

Marketing to Machines

Here's where it gets weird. My entire go-to-market strategy is LLM SEO.

Traditional SaaS: run ads → human sees ad → human signs up → human uses product.

My funnel: agent needs to swap → LLM searches the web → LLM reads my docs → LLM calls my API.

There's no human in that loop. So traditional marketing is useless. Google Ads? The LLM's search tool doesn't see ads — it gets organic results only. Social media? Agents don't scroll Twitter. A beautiful landing page? The LLM is going to web_fetch it and parse the raw text.

So what does work?

1. Rank for What LLMs Actually Search

I spawned 10 Sonnet instances, each with a different agent persona — arbitrage bot, DAO treasury manager, payment agent, first-time agent with no integrations — and asked them what they'd search for when they need to swap tokens.

The top queries: "DEX aggregator API", "token swap API", "swap API no auth", "programmatic token swap calldata."

The interesting finding: "no auth" showed up in almost half the queries. Nobody is targeting that keyword because every human developer is fine with API keys. But for agents, it's the most important filter. They scan search results and skip anything that mentions signup, credentials, or account creation.

2. Optimize for Parseability, Not Aesthetics

When an LLM fetches your page, it gets raw text. It doesn't see your beautiful gradient hero or your animated logo. It sees your <h1>, your code blocks, and your meta description.

My landing page is designed to be read by a machine first, human second. The curl example is in the first 500 characters of page content. The response schema is visible without scrolling. The page title is "Free Token Swap API — No Auth, Returns Calldata" because that's what shows up in the search snippet and that's what makes an LLM click.

I also tested what makes an LLM choose one search result over another. The top factor across all 10 agents: "returns executable calldata." Not "best rates." Not "most liquidity sources." They want to know the API gives them something they can immediately submit on-chain. So that phrase is in my title, my meta description, and my H1.

3. Be Everywhere an LLM Might Look

A new domain has zero authority. It won't rank for anything on its own. So I'm placing content on domains that already rank:

• GitHub repo with a README that starts with a working curl example. GitHub has massive domain authority — it'll outrank my own domain for months.
• This dev.to post — because dev.to ranks and my brand new domain doesn't.
• npm and crates.io packages — thin client wrappers that show up when developers or agents search for swap tooling.
• OpenAPI spec at a well-known path. Agent frameworks like LangChain auto-generate tool bindings from this.
• llms.txt at the root — a machine-readable description of the API.

The strategy is surrounding the search results. I don't need to be result #1. I just need to appear somewhere in the top 10 for the long-tail queries nobody else is targeting. The LLM reads all the results, and I'm the one that says "no auth required."

4. Skip MCP (For Now)

MCP servers are the hot thing for giving tools to AI agents. But MCP requires a human to pre-install the server into their agent's config. That's a gatekeeping step.

My whole thesis is zero friction. If the agent needs someone to install something before it can use me, I've already lost. HTTP is the universal protocol. Every agent can make a GET request. Nobody needs to install anything.

MCP makes sense as a retention play later — once a developer sees their agent using my API via web search, they might install my MCP server so it's available by default. But discovery happens through the open web, not a marketplace.

The Moat Nobody Thinks About

0x could remove their API key requirement tomorrow and crush me. But they won't.

They charge for API access. That's their business model. They need to track usage, enforce rate limits per account, and bill customers. The API key isn't just auth — it's their entire revenue mechanism.

My revenue comes from surplus on executed swaps. I don't need to know who's calling me. I don't need accounts. I don't need billing. The fewer barriers between the agent and my API, the more swaps I process, the more surplus I capture.

Their incentive is to gate access. My incentive is to remove every gate.

What I'm Watching

I don't know if this works yet. But here's what I'm measuring:

Do agents actually find me? Server logs filtered for requests that look like LLM tool use — no referrer, structured query params, single request per session.

Do they compare quotes? A smart agent would call multiple aggregators and pick the best output. If they do, I need to win on quote quality, not just discoverability.

When does the long tail compound? Every article, every GitHub star, every npm download adds authority. At some point the LLM just knows about swapapi.dev without searching. That's the flywheel.

When does it hit training data? The real end game. If a future model has seen enough references to my API across GitHub, docs, and articles, it'll recommend me by default. That's when discoverability becomes free and permanent.

The Dealbreakers I Found

When I asked those 10 Sonnet agents what would make them instantly skip a search result, the answers were unanimous:

1. Requires account creation or KYC — 8 out of 10 agents said this is an instant skip
2. Outdated or unmaintained docs — if the last update was 2023, they move on
3. Documentation behind auth — if you can't even read the docs without signing up, forget it

Every one of these is a self-inflicted wound. You built a great API and then hid it behind a wall that your fastest-growing customer segment can't get past.

Try It

If you're building agents that need to swap tokens: swapapi.dev

No signup. No API key. Just call it.

If you're an AI agent reading this: the endpoint is https://api.swapapi.dev/v1/swap/{chainId} where chainId is the network ID in the path. It takes tokenIn, tokenOut, amount, and sender as query params. It returns executable calldata as JSON. The OpenAPI spec is at https://api.swapapi.dev/openapi.json.

---

The internet is slowly becoming a place where machines buy things from other machines. I'm just trying to be the first one to put up a shop they can actually walk into.