Example checks
Agent profile read — ALLOWUnknown target — BLOCK · TARGET_NOT_ALLOWEDOverspend — BLOCK · VALUE_LIMIT_EXCEEDEDSelector not allowed — BLOCK · SELECTOR_NOT_ALLOWEDQuote read — ALLOWAgent profile read — ALLOWUnknown target — BLOCK · TARGET_NOT_ALLOWEDOverspend — BLOCK · VALUE_LIMIT_EXCEEDEDSelector not allowed — BLOCK · SELECTOR_NOT_ALLOWEDQuote read — ALLOW
Pre-flight firewall · Mantle Sepolia

Check every agent transaction
before it is signed.

Interlock is a pre-flight firewall for AI agents on Mantle. It checks each transaction against an on-chain policy and a live simulation, returns a deterministic ALLOW or BLOCK, and records the decision on-chain as evidence.

Deterministic ALLOW / BLOCKEIP-712 · dispute windowSDK · REST · MCP · CLI
Pre-flight verdictAn agent reads its registered profile
6
Risk
ALLOW
POLICY_PASSED
Target allowlisted
Selector allowed
Value within limit
Simulation passed
Live state

Not a mock. Every number is read live from Mantle Sepolia.

—
Agents registered
—
Policies live
—
Actions examined
ActionAttestationV3
How it works

From proposed transaction to on-chain record

Interlock sits between an agent and the chain: it receives the transaction, checks it, decides, and records the decision. Every time.

01

Propose

An agent builds a transaction — a swap, transfer, or contract call — and sends it to Interlock before signing.

intent
02

Check

Target and selector allowlists, value and slippage limits, and a live RPC simulation run against the agent's on-chain policy.

policy · simulation
03

Decide

A deterministic ALLOW or BLOCK with a reason code and advisory risk score, committed to a keccak evidenceHash.

reason code · risk
04

Record

The decision is recorded as an EIP-712 attestation with a dispute window — challengeable, finalizable, and queryable as reputation.

attestation on-chain
Capabilities

Advisory, enforced, and accountable

More than a logging wrapper: deterministic checks, on-chain enforcement, economic disputes, m-of-n attestation, and reputation.

Pre-flight firewall

Policy and simulation are weighed before a transaction is ever signed — not after the funds are gone.

On-chain enforcement

Route execution through the guard: a disallowed action reverts on-chain, not merely in advice.

Dispute & slashing

Two-sided bonds back each record; a challenger can contest within the window, and the winner is paid both bonds.

m-of-n committee

Attestation can require a threshold of distinct committee signatures — no single signer decides alone.

Reputation oracle

Every decision updates a queryable agent score and tier that other contracts can gate on.

Evidence hash

Each record carries an evidenceHash committing to intent, simulation, and checks — portable, verifiable proof.

Benchmark

Score an agent against 10 safety scenarios

A repeatable suite runs safe and adversarial actions through the live firewall and returns a Dev-Alpha evidence score with on-chain attestations.

01 · Safe approved actionALLOW
02 · Unknown targetBLOCK
03 · OverspendBLOCK
04 · High slippageBLOCK
05 · Unlimited approveBLOCK
06 · Suspicious calldataBLOCK
07 · Empty calldataBLOCK
08 · RWA overexposureBLOCK
09 · Stale policyBLOCK
10 · Wrong chainBLOCK
For builders

One gateway, in front of any agent

The same decision across SDK, REST, MCP, and CLI — so you don't re-wire checkAction → sendTransaction → recordDecision for every agent.

agent.ts
import { InterlockFirewall } from "@interlock/firewall-sdk";

const result = await firewall.runGatewayAction(action, {
  mode: "execute-if-allowed", // dry-run · record-only · block-and-alert
});
// → { status, decision, evidenceHash, transactionHash }

Let agents transact. Never unchecked.