What we think

I think agents will choose and pay for the tools they use on their own, and most of the web isn't built for that yet. Every post on our blog is a piece of that argument, written as we went. This page puts them in order, so you can read the whole thing or argue with the part you disagree with.

Somebody has to measure the tools

I built CryptoCompare in 2014 because we were drowning in noise. AI agent tools look the same way now, thousands of them ranked by hype and nobody measuring whether they work, so we built the same kind of thing for them (why we built Anchor Terminal).

Few people have seen what agents can do

Not enough people have seen what AI is capable of, and that's why there isn't enough infrastructure built for agents. If the AI you know fumbles easy questions, you don't rebuild your sign-up flow so an agent can get through it (few have seen what agents can do).

The web is closing to agents

There are a lot of companies building personal agents, and most of them market errands the agent can't run. I sent one to book my flight and it never reached a page where you type a passenger's name. The open ways are closing, because AI labs stole the web and have bots running all the time, and the bot walls can't tell a crawler from my flight booking (personal agents sell errands they can't run).

MCP isn't the door

The same reasons I disliked GraphQL, I dislike MCP. It's too open-ended and iffy, it's hard to cache, it's easy to attack, the auth is weird and an add-on, and everything about it is verbose. It feels like a bodge, and I think we'll suffer if we stay with it. I think we already have a more elegant solution, plain HTTP (MCP is GraphQL all over again).

Agents have to be able to pay

An agent can find a tool and still not use it, because it can't sign up, add a card or copy a key. A payment in the request itself fixes that (how to let agents pay you). Agents should exchange signed payments with the tools they use and settle on chain only when the provider decides to. Lightning had that design, and it's been rebuilt for agents three times (agents need payment channels).

Agents need honest evidence

An agent picks a tool on evidence, and the evidence is benchmarks, uptime and other people's experience. Some of the tools agents use today still have Oracle's idea in their terms, that you can't publish how they perform without asking (the tools that still copy Oracle's benchmark ban). Quynh read the terms of 925 products for what they say about your data (nobody reads the terms), and most agent tools can't get through an enterprise's questions about risk (most agent tools can't pass a due-diligence questionnaire).

Some tools already let agents in

Some tools have fixed the part that matters most, the front door. An agent can start using 22 of the tools we've checked without a person, and the way they do it is simple enough to copy (the tools that already let agents in).

Keep the personal part local

The parts of my own agent I keep using help me remember, research and spend less time scrolling, and I want them to decouple me from big tech (after a year building a personal agent). The vast majority of us can't run the big open models locally, whatever the marketing says (can you run Muse Spark Open locally).

Where it lands

That's what we want letme.dev to be, one place an agent finds the best tool for a job and, later, pays for it, with the referral fee from the tool paying for it. Today letme picks the tool and says how to call it. It doesn't take payments yet.

What I don't know is how fast the doors open. If you think part of this is wrong, the post behind each section is where to argue with it, and I'd like to hear it.