Execution firewall for AI agents.

Put a real execution boundary between AI and your tools.

Approve, refuse, or defer to human — before anything executes.
Every decision leaves proof.

AEG Gate is the public wedge for governed execution.
It shows a simple rule: AI can suggest anything, but it can only execute when AEG Gate issues one-time authority for that exact action.

Without AEG Gate:

agent -> tool -> execution

With AEG Gate:

agent -> AEK policy -> AEG token -> tool -> execution

const decision = await aeg.evaluate({
  action: "deploy",
  target: "service:smartlabel-web"
});

if (!decision.allowed) throw new Error("refused");

const token = await aeg.issueToken(decision.id);

await runDeploy({ token });

pnpm install
pnpm start:local
pnpm demo:10s

node ./bin/aeg.mjs run -- bash examples/protect-any-deploy/deploy.sh

1. A task is created
2. Execution without authority fails
3. AEK evaluates the request
4. AEG Gate issues a one-time execution token
5. Execution with the token succeeds
6. A proof artifact is emitted

Expected ending:

DEMO_10S_PASS
PROOF_PACKET_ID=...
PROOF_URL=...

Protected deploy ending:

DEPLOY_EXAMPLE_PASS

Release gate ending:

RELEASE_CHECK_PASS

This repo is the AEG Gate demo and open-source wedge.

It demonstrates:

• command intake
• task creation
• policy evaluation
• one-time execution token issuance
• token-gated execution
• proof artifact emission

pnpm start:local
pnpm smoke
pnpm demo:10s
pnpm prove
pnpm gate

This repo currently proves:

• token-gated execution
• refusal when no token is present
• proof emission
• local reproducible demo flow

This repo does not yet claim:

• multi-tenant SaaS
• public internet production hardening
• enterprise GA controls