ExchainExchain
DOCSREQUEST_A_PILOT →
HOME/ABOUT
WHO_WE_ARE

We don't fix
the AI.
We fix the
infrastructure.

Exchain is the control layer for agentic capital. Built for a world where autonomous agents move real money at machine speed — and where the math, not the model, must be the last line of defense.

HARDWARE ENCLAVE NETWORK · LIVE
§
TRUST_DOMAINS
3
Decision · Custody · Settlement
ENCLAVE_QUORUM
2/2
Attested hardware co-sign
KEYS_HELD_BY_AGENT
0
It proposes. It never signs.
HUMAN_APPROVALS
$0
Math is the approver
01 / ASSUMPTIONS

// Our operating assumptions about agentic finance

01The AI will hallucinate.
02The AI will be prompt-injected.
03The AI will execute blindly.
04We assume all of it.
05We build for failure.
06Not hope.
02 / NOT_THIS

What Exchain
is not.

The agentic economy needs new primitives — not wrappers around existing ones. Hover each to understand the distinction.

✕
A probabilistic AI guardrail01

Guardrails guess if the AI is thinking correctly. When the model breaks, the guardrail breaks with it.

✕
A blind execution pipe02

x402 and streaming payment APIs are gas pedals. We are the brakes. They are not the same infrastructure.

✕
An access-based custody wallet03

Custody asks who holds the key. We ask whether the intent is mathematically legal. Different question entirely.

✕
An LLM fix04

We do not touch the probabilistic layer. We intercept the intents at the infrastructure level before they hit the execution pipe.

THE_RIGHT_QUESTION
Custody asks:
"Who holds the key?"
Guardrails ask:
"Is the AI thinking correctly?"
Exchain asks:
"Is this intent mathematically legal?"
03 / MECHANICS

Machine speed.
Mathematical enforcement.

One agent. One constitution. Zero keys. Here is what happens to every intent.

DECLARE
01
Write the constitution
Budgets, rolling spend windows, per-counterparty caps, velocity limits — declared in a deterministic policy language and compiled to stateful invariants. Your CFO’s numbers, made executable.
→
PROPOSE
02
The agent emits an intent
The agent holds no keys. It proposes a spend — a payment, a hire, a trade — and the intent hits the Exchain gate before any payment rail ever sees it.
→
EVALUATE
03
Two independent judges
The policy engine evaluates the intent against the live state of every invariant. Attested hardware enclaves re-evaluate it independently and co-sign. Disagreement is a denial.
→
ALLOW or DENY
04
Math decides
Allowed intents execute at machine speed, carrying the enclaves’ joint signature in the journal. Everything else fails closed — and the worst case stays under the ceiling you set.
TRUST_ANCHORS:PYTH_NETWORKCHAINLINKEASAWS_NITRO_ENCLAVES
04 / PRINCIPLES

Zero trust
architecture.

PARA_01
01
Assume total failure
The intelligence layer will fail completely. Every system is built expecting hallucination, injection, and a compromised host — simultaneously.
PARA_02
02
Distrust is the product
Nothing rests on trusting us. The binaries run on your box, the keys stay in hardware, and every claim ships with the script that lets you verify it yourself.
PARA_03
03
Deterministic constitutions
Hardcoded math replaces probabilistic guesses. If it cannot be proven, it cannot be trusted.
PARA_04
04
Fail-closed denial
A violating intent is denied before money moves. Not a warning. Not a log. A denial — and ambiguity counts as a violation.
PARA_05
05
Judge, not custodian
Exchain decides whether money may move. It never holds the money, and it never holds your keys.
PARA_06
06
Default alive
Capital remains intact when the AI breaks. The system degrades toward no spending — never toward unbounded spending.
05 / POSITIONING
We don't fix the AI.
We secure the capital mathematically.

Exchain is the layer the agentic economy stands on. We intercept agent intents before they reach the payment rail, enforce deterministic constitutions across independent trust domains, and bound the worst case in math. When the AI breaks — and it will — the most you can lose is a number you chose. Not even Exchain can exceed it.