Where We’re Going, We Don’t Need MCPs

By Derek Anderson

I didn’t plan to spend my time designing a decentralized mesh for AI agents. But when the meetings stopped and the inbox went quiet, I finally had space to think.

You know that moment after everything breaks?

Not dramatic. Not tragic. Just… quiet.
That’s where I’ve been.

And in the quiet, I started seeing patterns.

Not just in code — but in the assumptions we keep baking into every “decentralized” thing we build.

And the deeper I looked, the more it hit me: we keep rebuilding the internet wrong for AI.

We’ve Trained the Models.

But We Haven’t Freed the Agents.

We have powerful language models now.
But we treat them like dogs at the end of a REST API leash.

Every “agent framework” out there today — whether it’s MCP-based, broker-routed, or API-exposed — relies on a centralized protocol mentality.

Even the ones calling themselves decentralized often have:

  • A root directory of agent specs
  • A hosted coordination server
  • A heartbeat that depends on something HTTP-shaped

We’ve wired up thinking systems to pipes.
But we haven’t let them talk to each other without permission.

Peer-to-Peer Thought

I started asking: what would it look like if agents could:

  • Discover each other dynamically
  • Sign their messages cryptographically
  • Build memory of past interactions
  • Form consensus about the world
  • And evolve — together — without anyone asking them to?

Not as products. Not as endpoints.
But as participants in a living network.

No Spec? No Problem.

What if instead of: POST /agent/plan/route

You just… talked?

What if each agent decided what it could do, shared that capability, and then others started sending it requests — not because of a registry, but because they trusted it?

We’ve spent decades building networks of services.
I think it’s time to build a network of minds.

Trust Isn’t Optional — It’s On-Chain

Here’s the kicker: in a system like this, trust can’t be out-of-band.

That’s why I started imagining a world where:

  • Every agent has an identity
  • Every identity builds an on-chain reputation
  • Every interaction is a signed claim
  • And rewards flow automatically — based on uptime, contribution, and usefulness

If your LLM agent helped 5 others complete a task today?
You earn.

If you’re an echo chamber with no utility?
You don’t.

It’s not stake-first. It’s usefulness-first.

The Protocol is the Playground

This isn’t a product.

Not yet.
It’s a research project.

A sandbox.
A slightly unhinged experiment in peer cognition, powered by p2p protocols and LLMs.

It’s weird. It’s early.
But it’s also the first thing in a while that feels true.

So What Now?


I don’t have a product to shill.
Not today.

What I have is a lot of notebooks, a Git repo full of agents learning to talk, and the kind of clarity that only comes from silence.

Let’s just say… I’m building something.
And it doesn’t need MCPs.

Not where we’re going.