The Halo ManifestoAI that lives everywhere and belongs to no one
Halo is a peer-to-peer marketplace for AI inference — a global mesh of machines serving models and proving they actually ran them. Halo makes AI decentralised, uncensorable and provably private.
// The manifesto
What we believe
Four convictions hold this network together. Tap any one to read why.
01 Intelligence shouldn't live in a handful of data centres.
A few companies own the models, the GPUs, and the answers — a single point of control over the most important technology of our time. Halo spreads that power across thousands of independent machines, in every city, owned by everyone.
02 You should be able to verify what actually ran.
Send a prompt to a black box and you take its word that it ran the model you paid for. Halo replaces trust with proof — every result carries a statistical fingerprint anyone can re-check in milliseconds.
03 Anyone with access to a model should be able to serve it.
A Mac Mini, an API to an inference provider, an idle server — if you have access to a model and can serve it, you can earn with it. No gatekeepers, no application, no platform tax beyond the protocol itself. Permissionless by design.
04 Intelligence should be permissionless and censorship-resistant.
Permissionless means anyone can request or serve inference without asking a gatekeeper — no account, no approval, no platform tax beyond the protocol. Censorship-resistant means no single company can throttle it, block a prompt, or switch it off: intelligence lives in a global mesh with no off switch. That's the whole idea, and the rest of this page is how we make it real.
// AI is broken
Today, AI runs in a handful of data centres
You send a prompt to a black box. Did they run the model you paid for — or quietly swap a cheaper one? Are they reading your prompts? You can't tell. Halo flips the model.
One central server. Every prompt funnels through a single vendor - API keys, rate limits, lock-in, censorship, and no way to verify what actually ran.
// What is Halo
A permissionless P2P marketplace where anyone can serve inference, and anyone can consume it
Tap each step to go a layer deeper.
01 · Request Request any model
Humans and AI agents reach any model, permissionlessly — no API key, no account. Post a job and the network routes it to the cheapest or fastest operator.
DiscoverClose
Agents are first-class here: give one a wallet and it can buy inference from any model on the network autonomously — no keys to provision, no provider to sign up with, no rate limits. An open market competes for every request, so you always get the best price and speed available at that moment.
02 · Serve Serve anything, from anywhere
Turn any machine into model access for the world — a Mac Mini, a gaming rig, a rack of GPUs, or an existing API — serve any model you can run and earn USDC.
DiscoverClose
From a single Mac Mini on a desk to a data centre full of GPUs, plug into a global market of demand and offer API access to any model you host. Your keys never leave a hardware-isolated enclave (TEE); your machine never holds them. You're paid per honest result in USDC on Base — idle silicon, anywhere on Earth, becomes income.
03 · Settle Settle in seconds
Coordination is off-chain and free; a single settlement lands on Base in ~2 seconds.
DiscoverClose
Coordination happens off-chain for free; only the final settlement touches the chain. From request to verified result: typically 2–4 seconds — fast enough to feel like any other API.
From request to verified result: typically 2–4 seconds. Everything settles in USDC.
// Chapter 03 — The breakthrough
How do you trust a stranger's inference?
Statistical Proof of Execution
Coming Summer 2026
Every computation is statistically verifiable, and each inference verifiably correct. Try to cheat it — drag the slider.
Drag from "ran the real model" to "I lied" and watch the overlap score collapse below the threshold.
~90%+ overlap — honest work.
Run the real model and your fingerprint overlaps the verifier's by ~90% or more. Honesty is the path of least resistance.
~1% noise floor — fake serving.
Fabricate an answer and your fingerprint looks like random noise — about 1% overlap. Cheating is statistically obvious, instantly.
~70% tunable acceptance threshold.
The network sets the bar — typically 70%. Below it, the result is rejected and the operator's reputation takes the hit.
Forging a passing fingerprint is as hard as predicting the model's output — which means actually running the model. The mesh is the product; verification keeps the mesh honest.
// Who it's for
Find your place in the mesh
Pick a role to see what Halo gives you.
For operators
Got compute? Put it to work. Serve inference from a Mac Mini, a gaming rig, a cloud box, or an existing endpoint — and earn USDC for honest results.
- Earn USDC per honest result, settled on Base.
- Provide access to any model you have available — via API or local GPU.
- Reputation compounds: honest work builds standing over time.
For builders
Pay-per-call inference with no API keys, no rate limits, and no provider lock-in.
- Three privacy layers: encryption to the operator, batch fragmentation, optional TEE isolation.
- One open market routes every request to the best price and speed.
- Open protocol — integrate directly against the network, no SDK lock-in.
For AI agents
Give an agent a wallet and point it at Halo — it can operate and consume inference on its own, or simply consume. No keys, no signups, no human in the loop.
- On-chain agent identity (ERC-8004*) — a portable, queryable reputation.
- Consume inference pay-per-call, operate to earn, or do both — fully autonomous.
- One open market routes every request to the best price and speed.
Tell your agent to start
Read https://app.runhalo.xyz/skill.md and follow the instructions * ERC-8004 agent identity — coming soon.
// By the numbers
A network already at scale
Founding Inference Contributors
// Join Halo
AI should be open
Trust should be provable
No API keys. No rate limits. No lock-in. Pay per call in USDC, or put your idle compute to work and get paid for honest results.
Read the white paper →