Pre-action authority

Govern proposed agent actions before execution

Neura Relay reviews proposed consequential actions before real systems change. It returns a Decision Receipt with allow, review, revise, or stop before execution happens, while your runtime stays in control.

Run authority flow Open refund example

What Relay does

Before action: Risk checked first
Receipt first: Receipt recorded
Runtime control: Authority checked

Authority path

Agent intent becomes a receipt before action

Agent wants to act

Runtime prepares the proposed actionbefore anything downstream changes

Action Card

Intent, actor, target, and policyevidence and risk are made explicit

Authority review

Relay classifies the proposed actionand records the authority decision

Decision Receipt

System reads the receipt outcomethen routes execution downstream

Sequence-aware routing

Relay reviews the next proposed action with the steps that led to it

Workflow sequence

The agent moves step by step

Every proposed action carries recent workflow contextfrom the steps immediately before the next change

Recent context

The next action is reviewed

Relay reviews the sequence before the next changeso the route reflects what just happened before

Receipt route

The runtime gets a route

The runtime receives a route it can act oncontinue, revise, review, or hold with the reason

Routing result: the same next action can route differently when the preceding sequence changes the risk, while execution stays in the runtime

Authority examples

Five local examples, one Relay mechanism

Action firewall

Classify risk firstbefore systems change

Receipt standard

Record the decisionwith inspectable refs

MCP risk gate

Bind tool and paramsbefore execution

Authority path

Check authority pathand sequence context

CommerceOps drill

Stress-test refundswith synthetic data

Run locally

Clone the local examples and inspect the receipts

Run the local authority flow to see exactly how Relay pre-action authority works before any real system is touched. It creates receipts for proposed actions, shows why each action is allowed or held, and keeps execution in your runtime.

Run the authority flow

git clone https://github.com/neurarelay/relay-action-card.git
cd relay-action-card
npm ci
npm run verify:decision-receipt-standard
node scripts/run-agent-action-gateway.mjs --dry-run --json

Refund approval example

A synthetic $420 refund shows the route for money-moving actions

Proposed

$420 refund

Agent proposes refund and customer notice

Decision

Owner review

Relay pauses the action before anything changes

Blocked

Payment held

No refund or customer message is sent

Open refund example

Example amount only; no real store, payment, or customer message

Boundaries

The product boundary is deliberately narrow

Relay is the product; authority is the surface shown here
Action Cards go in; Decision Receipts come out
Your runtime still executes, revises, reviews, or stops

Authority checks are receipt factors, not certification claims
Examples use synthetic data and do not touch real systems
No provider approval, public tokens, Registry auto-approval, or Neura execution