The AI Runtime Control Plane

Find your AI. Protect it. Prove it.

Norcaster finds AI apps your company already runs, puts them behind one control layer, and keeps a record of what they did.

Run a policy checkBook a 20-min demo
  • EU-hosted or self-hosted
  • OpenAI- & Anthropic-compatible
  • Independently verifiable evidence
Your AI. One control plane.Architecture overview

01 / Connect

AI applicationsYour product
Agents & copilotsYour assistants
WorkflowsYour AI processes

OpenAI- & Anthropic-compatible APIs

02 / Control

Norcaster

AI Runtime Control Plane

Deployment & policy version
Policy enforcement
PII redaction
Budget controls

Enforce before inference

03 / Execute

Models & providers

OpenAI
Anthropic
Google
Groq
Together
Bedrock
VoyageEmbeddings
Your own models

Observe & evaluateConnected by request identity

Traces & costQuality evalsVerifiable evidence

Controls apply on supported production paths, as configured.

Provider integrations

  • OpenAI
  • Anthropic
  • Google
  • Groq
  • Together
  • Bedrock
  • Voyage
View integrations →

01Today, without Norcaster

Every team connects to AI on its own

AI spreads across teams, tools and providers. Without a shared control layer, visibility, policy and evidence stay fragmented.

The fragmented AI estate

Illustrative setup

No shared control layer between your applications and AI providers.

  • Fragmented visibility

    AI usage is spread across teams, tools and provider accounts.

  • Inconsistent controls

    Each application follows its own rules instead of one shared policy.

  • Scattered evidence

    Reviewers have to piece together what happened from separate systems.

See what changes with Norcaster

02With Norcaster

One control layer between your apps and AI

Bring your AI requests through one runtime control plane. Apply policy before the model runs and keep a record of each decision.

Your AI, connected through Norcaster

Architecture overview

Controls apply on supported request paths, as configured.

  • Centralized visibility

    See requests and decisions for AI routed through Norcaster.

  • Consistent controls

    Apply shared rules before requests reach the model.

  • Traceable decisions

    Keep the policy version, rule and timestamp with each decision.

Start with one application

For each app it is a small change: it sends its AI requests to Norcaster instead of straight to the provider.

03How it works

Three steps

  1. Find

    See AI apps you already run.

    A small scanner runs on your own systems. It reads your code and network logs and lists the AI apps it finds.

    Your code stays with you. Only names and counts are shared.

    Shared

    • App and repo names
    • AI providers and models
    • Counts and setting names

    Never shared

    • Source code
    • File paths and line numbers
    • Key values

  2. Protect

    Put each one behind Norcaster in a few clicks.

    Pick the app. Accept the suggested rules. Confirm. Norcaster sets it up and writes the code change for your developers to review.

    Nothing changes without a person saying yes.Example

    Customer chatbot

    Suggested rules
    Enterprise baseline pack
    Approved by
    A person on your team
    Code change
    Ready for your developers to review

  3. Check

    See which apps are protected, and close the gaps.

    For apps that still go around Norcaster, your IT team gets ready-made blocking rules. Then we show you whether that traffic stopped.

    Your AI appsExample
    • Customer chatbotProtected
    • Card-dispute summariserStill goes around Norcaster
    • Payments agentNot connected yet

    Protected means Norcaster saw that app's requests come through in the last 24 hours.

04The result

What you get

  • One overview

    Every connected AI app, who used it and what happened.

    Connected AI appsExample
    AppRequests, 24 hBlocked or masked
    Customer chatbot1,28437
    Back-office tool3124
    Developer tools4,906112

  • Sensitive data stopped

    Personal data and secrets are blocked or masked before they reach the AI.

    What the AI receivesExample

    Your app sent

    Please draft a reply to j.smith@example.com about the refund for order 4471.

    The AI received

    Please draft a reply to [REDACTED: EMAIL] about the refund for order 4471.

  • Evidence for your auditor

    A tamper-evident record of every decision, ready to export. Included from Growth up.

    Evidence recordExample

    Customer chatbot · production

    Decision
    Allowed, email address masked
    Policy
    Enterprise baseline pack · 1.2.0
    Detected
    Email address
    Action
    Masked before the AI provider
    Provider
    OpenAI · gpt-4o
    Recorded
    30 Sep 2026, 14:02:11 UTC

    Exportable for your auditor · tamper-evident from Growth up

Governed requests leave a defensible record — policy decision, redactions, provider, timestamp, trace ID — exportable as an audit bundle. Norcaster produces the evidence; your advisor renders the judgment. Norcaster does not declare anyone "compliant." Seven evidence artifacts — how proof works →

05Agents

Its own identity, its own limits, and a way to stop it

Register each agent in your organisation. It acts under its own name, inside the limits you set, and you can stop it with a record of who did it and why.

  • Each registered agent has its own identity

    Register an agent and it gets its own keys, scopes and deployment bindings. The records of its model calls and tool calls name the agent, with its owner.

    AgentExample

    refund-agent

    Owner
    Payments operations lead
    Keys
    1 active · scoped to runs
    Bound to
    Customer refunds · production
    Request rate
    60 a minute

    Its decision records name the agent and its owner

  • Limits on what it may do

    Its keys are scoped and bound to the deployments it may call, with its own request-rate limit. Each tool call is allowed, redacted, blocked, or paused for approval by a person other than the one who asked; unregistered tools are blocked by default and irreversible actions fail closed.

    Tool rules, from its deploymentExample
    ToolOutcome
    get_orderAllowed
    issue_refundPaused for a person
    close_accountBlocked
    Any unregistered toolBlocked

  • A kill switch with an audit trail

    Retire one agent and all its keys are revoked at once. Stop one deployment, or engage the kill switch to refuse all runtime traffic for your organisation. Each stop is recorded with who, when and why, and every refused request is a blocked decision on the evidence record.

    Audit trailExample
    18:04 · Deployment stopped
    On-call engineer · “Refund loop in the 18:00 batch”
    18:04 · Request refused
    Blocked decision on the evidence record
    18:11 · Agent retired
    Payments operations lead · all keys revoked

A stop refuses new requests on the next call, on every API replica. Responses already streaming end within about two seconds; a running step finishes. Hybrid data planes (Sync Agent installs) do not receive the stop yet. Anyone who can deploy to a deployment can stop it; resuming it, retiring an agent and the organisation kill switch are for owners and admins. Agents in the docs → How the kill switch works →

06Shared Record

In development · pilot pairs by application

Next: one record both companies can check

When your AI agents deal with another company's agents, Shared Record will give both sides one record of the exchange, signed by both and checkable by anyone offline.

How Shared Record works
Illustrative exchange
Company AAI agent
Company BAI agent
Shared Record

One exchange.
Two signatures.

Signed receiptCompany A
Signed acknowledgementCompany B
Shared logFingerprints, no content
Outside verifierChecks the export offline

How it works in the pilot version

  1. Send with a signed receipt

  2. Partner signs an acknowledgement

  3. Fingerprints on a shared log

  4. Anyone checks it offline

07Built for regulated companies

Safe to start with

  • Hosted in the EU

    Nuremberg, Germany.

    Operated by
    Navego AB, Sweden
    Subprocessors
    Published list →

  • Or in your own environment

    Run it yourself if you prefer.

    Licence
    Checked locally
    Calls to Norcaster
    None, unless you turn on update checks

  • Start in watch-only mode

    Record first, block later.

    Last 7 days, watch onlyExample

    Requests recorded
    9,840
    Personal data found
    52

    Set per app. Switch when you are ready.

    Each pack's public manifest lists what stays blocked even in watch-only mode. See the manifests →

Security posture

  • SOC 2 Type II: Window opens Nov 2026
  • ISO 27001: On roadmap
  • ISO 42001: Planned
  • Third-party penetration test: Planned · in Type II window
  • Evidence integrity: Append-only · Rekor
  • Deployment: EU-hosted · self-hosted
Trust center →

08Become a Norcaster partner

The AI Runtime Control Plane

Your expertise.Our runtime evidence.

For consultants, advisors and firms who help companies run AI in production under rules they have to show they follow.

Explore the partner program
01

Independent

Consultants, advisors, fractional CISOs and lawyers who work on their own.

Introduce clients who need controls and runtime evidence for the AI they run.

Apply as an independent
02

Company

GRC and advisory firms, managed service providers, resellers and technology vendors.

Advise, sell together, resell, or build an integration with Norcaster.

Apply as a company

Partners advise. Norcaster produces the runtime evidence. Auditors and regulators decide.

We read every application ourselves and reply within two business days.

The AI Runtime Control Plane

Start with one governed AI request.

Run a policy check now, then see how the same control plane deploys, controls, observes, and evaluates production AI.

  • Try it now

    Paste a made-up prompt and see what a real policy pack does with it.

    • No signup
    • Nothing stored
    • No AI provider called
    Run a policy check

  • Talk to us

    A 20-minute walkthrough of finding, protecting and proving, with time for your questions.

    • Pick a time online
    • Bring one AI app you run
    Book a 20-min demo

  • Security review

    What your security team usually asks for, in one place.

    • Security posture, stated plainly
    • Subprocessor list
    • Published detector manifests
    Visit the Trust center