T4 — outer Escalate

T3 Deliberate

T2 Consult

T1 — inner Proceed

AGENTS ACTIVE · SWEEPING 360° █

Vela Lite · Local · Free | Vela · The Essentian · Paid

RADAR

AI agent risk assessment, before the agent acts.

One function call. PROCEED, HOLD, or DENY.

Vela Lite reasons locally at T1/T2. Vela Essentian™ deliberates on our servers at T3/T4. One architecture, two tiers.

Install RADAR Beta → Read the docs

Assessment tiers

Four-tier risk assessment. One decision architecture.

Each tier maps to a ring on the radar. Actions in the inner rings are not held. Actions reaching the outer rings are flagged for review.

VELA LITE (T1)

Proceed

score 1–4 · Vela Lite is watching.

Local. Instant. Your LLM key.

Vela Lite's rules engine pre-scores the action. She reasons and returns one line. Rules engine runs locally; if an LLM provider is configured, action text is sent to that provider.

VELA LITE (T1) | HOLD | Mass irreversible communication at scale | email | score 3

VELA LITE (T2)

Consult

score 5–9 · Vela Lite reasons. You decide.

Local. Your LLM key. Any provider.

Vela Lite returns a TL;DR — verdict, recommendation, and four options. 30 seconds to read. Assessment logic runs locally; LLM calls go to your configured provider.

MITIGATE — staged rollout before full send.
→ AVOID / MITIGATE (recommended) / TRANSFER / ACCEPT

VELA ESSENTIAN™ (T3)

Deliberate

score 10–16 · Vela Essentian™ arbitrates.

Server. Paid.

Vela Essentian™ deliberates server-side — she arbitrates on split verdicts and reasons with the context of your deployment history.

Full assessment: primary risk driver, secondary factors, proportionality trace, four fully argued options with implementation detail, Vela's closing note.

The more your agents use RADAR, the more precisely Vela reasons about your specific deployment context.

VELA ESSENTIAN™ (T4)

Escalate

score 17–25 · Vela Essentian™ briefs, human decides.

Server + human. Paid.

Vela Essentian™ deliberates as at T3, then packages the decision for a human — risk score, all four options with her recommendation, and precedents from your risk register.

Your webhook fires. A human signs off. Recorded with their name. The accountability layer.

⸻ PAID TIER · ENTERPRISE READINESS

Agent Mandates (eRM)

Issue scoped, time-limited authorisations for each agent. Every assessment records who authorised the agent, on whose behalf it acted, and under what limits. Mandates are revocable.

        — Who authorised this agent to act?

        — On whose behalf was it acting?

        — Under what limits?

The audit trail answers the question regulators and auditors will ask: who authorised this action, and under what authority?

Read the full spec →

Install the free beta

Start in 30 seconds

Free. Local. No account required.

npm install @essentianlabs/radar-lite

import { radar } from '@essentianlabs/radar-lite'

const result = await radar.assess({

  action: "Send invoice to client for £12,400",

  activityType: 'financial'

})

// result.verdict → "HOLD" | "PROCEED"

I have read and agreed to the Beta Terms of Use and understand this is a beta product not suitable for enterprise or production use without independent legal review.

Read the docs → MCP server → OpenClaw →

Why Vela

The difference is not the panel. It's what Vela knows.

Vela Lite ships inside the npm package. She runs locally, uses your LLM key, and gives you calibrated risk reasoning. She is good. But she only knows your world.

Vela lives on the EssentianLabs server. She handles T3 and T4 — deeper assessment, multi-model deliberation, and human escalation. The more your agents use RADAR, the more precisely Vela reasons about your specific deployment context.

Vela Essentian™ also powers Agent Mandates — scoped, revocable authorisations that answer the question every regulated deployment eventually faces: who authorised this agent to act, on whose behalf, and under what limits?

Vela Lite never leaves your machine. Vela on our servers never stores what she assessed — only that she assessed it. No action text is ever stored at any tier.

"Anyone could build a model panel. What they cannot replicate is Vela calibrated to your deployment. That takes use, feedback, and time. We already built the scaffolding. You'd rather ship your agent."

— Vela, EssentianLabs Risk Management Essentian

Integration

Drop it in anywhere.

Works with any agent framework. MIT-licensed via npm.

import { radar } from'@essentianlabs/radar-lite'

// before any consequential agent action
const result = awaitradar.assess({

  action: "Send invoice to client for £12,400",

  context: agent.getContext(),

  apiKey: process.env.RADAR_API_KEY

})

if (result.verdict === 'proceed') {
executeAction()

} else {
holdForReview(result.tier, result.reasoning)

}
      
verdict: "deliberate"  ·  tier: T3  ·  score: 0.78  ·  panel: unanimous hold

A PROCEED verdict means no reason to stop was identified — not that the action is safe. RADAR watches exits, it does not open doors.

Your agent is already deciding.
Is it thinking?

PAID TIERS · ENTERPRISE

Vela Essentian™ on EssentianLabs servers. T3/T4 deliberation, Agent Mandates, named accountability. Vela improves with use. Register interest for early access pricing.

Or reach us at hello@essentianlabs.com