Platform

One control plane for every agent you run.

Sentinel NHI sits between your AI agents and the tools they use. It knows who each agent is, decides whether each action may run, brings a person in when it matters, watches for patterns no single call reveals, and keeps a record nobody can quietly edit.

See a live demo How it integrates

Identity and ownership

Every agent has a name, a purpose and a person

  • A registry of agents with risk tier, environment, allowed models and an accountable owner (and backup owner)
  • Unknown agents that call in are recorded as discovered, never trusted by default
  • Discovery of other non-human identities: service accounts, keys and tokens across your cloud and code platforms
  • Findings for risky combinations, such as an agent that can read private data, take in untrusted content and send data out

Credentials and delegation

Short-lived, signed, and only ever narrower

  • Agent tokens live 15 minutes at most and are signed with keys held in Azure Key Vault, AWS KMS, Google Cloud KMS or HashiCorp Vault, including HSM-backed keys
  • Sub-agents get delegated tokens that can only narrow tools, limits and depth; revoking a parent revokes its children
  • Workload identity federation: agents on Kubernetes or AWS authenticate without a stored secret
  • Revocations are pushed to enforcement points in seconds

Authorization

Every tool call decided before it runs

  • Allowlisted MCP servers and tools, argument constraints, rate limits, data scopes and spending budgets
  • Tools are classified by what they do; payouts, exports and deletions are irreversible and need a person by default
  • Just-in-time grants for the rare task that needs more, with expiry
  • Sentinel's MCP proxy authorizes every call without changing agent code

Human oversight

People decide where it matters

  • An oversight matrix by tool and risk tier: run, notify, approve, co-pilot (a person edits the arguments first) or deny
  • Approvals in the console, Slack or Microsoft Teams, with escalation to on-call approvers and expiry into a deny
  • Containments narrow an agent to read-only; the kill switch pauses it or quarantines a session at once
  • Kill switch drills prove the stop works before you need it

Monitoring

The pattern, not just the call

  • Correlation rules for bursts with kill-chain progression, credential harvesting, enumeration, audit tampering and unmanifested compute
  • Input-to-action provenance: untrusted input, then a private read, then data leaving, in one session
  • Honey tokens, persistent-denial alerts with pause proposals, parity testing and shadow mode
  • Signals from CipherWatchSuite Guardrails (prompt injection, data leakage) narrow or stop an agent

How security teams use it

Audit and evidence

A record no one can quietly edit

  • Every decision, approval and change is appended to a per-tenant hash chain before the action runs
  • Signed checkpoints, an optional external witness, and RFC 3161 trusted timestamps
  • A forensic timeline from the human principal down through every sub-agent
  • Continuous controls and signed evidence bundles mapped to SOC 2, ISO/IEC 42001, NIST AI RMF and the EU AI Act

Built for teams

Multi-tenant, role-based, self-hosted.

Tenants

Business units or customers are separate tenants, isolated in the database with row-level security.

Roles

Admin, policy editor, approver, auditor and agent owner, each with its own permissions.

Single sign-on

OpenID Connect sign-in, with group-to-role mapping.

Your cloud

Docker or Kubernetes; your database, your key vault, your network.

See Sentinel govern a real agent fleet.

A live demo of a fictional fintech running ten agents: approvals, a containment, an attack caught in the pattern, and the evidence for your auditor.

Book a demo See plans