Marquor Runtime Security - Enterprise

Marquor Enterprise is the licensed, capacity-aware edition of Marquor's enforcement and evidence layer for AI agents. Every consequential action an agent takes, allowed and denied alike, becomes an Ed25519-signed, hash-chained receipt that verifies offline, with no call back to Marquor. Because the chain is hashed, a retroactive edit or deletion is detectable. A third party can check your evidence without trusting us.

Each receipt records what a given agent was permitted to do, what it did, and on whose delegated authority. That answers the question which arrives long after the action: not whether you have logs, but whose word your records rest on.

• Logs are a system's account of itself.
• A registry records design-time intent and does not constrain a running agent.
• A periodic audit speaks to a control environment over a sample, not to this action on this day.

Marquor produces the per-action record none of those can, and you keep it as your own compliance artifact: producible to a regulator, a customer, an auditor or your insurer at renewal, without Marquor in the loop.

Marquor evaluates declared metadata only. Prompts, files and transaction bodies are never collected: spend bands rather than amounts, digests rather than payees. The decision itself is deterministic, made in real time against verified identity and signed policy, no LLM sits in the decision path, so the same inputs give the same answer, in microseconds on CPU.

Because enforcement sits at the agent-to-tool boundary, Marquor governs agents whichever model or framework they are built on, and complements your existing identity layer rather than replacing it. You do not have to standardise your agent stack first. One control plane covers every enforcement point you deploy, whether that is an Amazon EKS cluster, EKS Anywhere, or a self-managed Kubernetes cluster in an isolated VPC.

Enforcement is something you switch on, not something that happens to you: every policy can run in monitor mode first, so you see the decisions before any of them take effect.

• Entitlement-aware capacity. The gateway verifies your AWS Marketplace subscription through a non-consuming license read that can never draw down purchased quantities. Edition and enforcement-point capacity are discovered from the granted license, so a subscription change reaches running gateways on their next check with no redeploy.
• Licensing that fails safe. A bounded grace window keeps an already-verified gateway serving straight through a transient AWS outage. Until a subscription has been verified at all, the gateway declines rather than serving unlicensed traffic.
• Egress to destinations you approve. Outbound calls go to the destinations on your allowlist; anything else is refused and written to the same evidence ledger as tool-call decisions, as a signed observation. Start in monitor mode to see exactly what your agents already reach, then enforce when the list looks right.
• Editions and add-ons. Launch, Platform and Complete cover 25, 75 and 150 governed agents with 3, 6 and 12 enforcement points. Enforcement-point packs, additional isolated tenants, compliance-framework crosswalks with signed evidence bundles, a self-hosted or VPC-isolated sealed engine, and a named technical account manager with a 99.95% availability SLA are available as separate dimensions.

The ledger is single-writer and the chart refuses to render above one replica. Configuration loaders fail safe: the gateway will not boot on an invalid document, and a caller whose authority it cannot verify is refused rather than admitted on assumption.

Runs on Amazon EKS, EKS Anywhere and self-managed Kubernetes, multi-architecture on x86-64 and AWS Graviton, as a non-root user on a minimal base image. Marquor has passed the AWS Foundational Technical Review and completed an AWS Well-Architected Framework Review.