Concept
AI gateway vs AI runtime control plane
An AI gateway is a proxy in front of model providers: applications point their SDK at one base URL and the gateway handles routing, credentials, retries, and usage metering. An AI runtime control plane sits in the same position but treats each request as a governed decision — policy is enforced before the model runs, the outcome is recorded as evidence, spend is gated, and quality is checked — with a control surface for configuring all of it. Norcaster is a runtime control plane that includes an OpenAI- and Anthropic-compatible gateway as its ingress.
What an AI gateway does well
- One base URL and one credential model in front of several model providers.
- Routing, retries, and failover between providers without changing application code.
- Usage metering per key, team, or application.
- Request logs and latency metrics for operations.
These are real problems, and a control plane has to solve them too. The difference is what happens to a request once it has been routed.
What a runtime control plane adds
| Capability | AI gateway | AI runtime control plane |
|---|---|---|
| Where it sits | In front of model providers; applications change their base URL. | The same position in the request path, plus a management surface that decides what the runtime enforces and lets you inspect what it did. |
| Provider routing and failover | Routes across providers with retries and key management. | Routes across providers, and records each failover as an event on the request trace. |
| Policy before the model runs | Typically rate limits, key scopes, and simple content filters. | Every request is allowed, blocked, redacted, or budget-gated against a versioned policy — schema contracts, contextual detectors for personal data and adversarial input, and enforcement packs — before the provider sees it. |
| Evidence per request | Request logs and metrics you assemble yourself. | An enforcement record per request: policy version in force, rules evaluated, verdict, and reason. Redactions are recorded as masked metadata, never the sensitive text. Traces are hash-chained and anchored to a public transparency log. |
| Spend control | Usage metering; sometimes a budget per key. | Budget gates that refuse a request before it runs, with cost and latency captured on the trace. |
| Quality gates | Out of scope. | Evals tied to deployments and traces, so a candidate that regresses below threshold can be blocked. |
| Audit deliverable | Logs exported to your own tooling. | Framework-scoped audit bundles (EU AI Act, SOC 2 Trust Services Criteria) that verify with a standalone script — no vendor account needed. |
The gateway column describes the category in general terms; check a specific product’s documentation for what it ships.
When a gateway is enough
- You need one API in front of several providers and nothing in the request may be blocked or altered.
- No regulator, auditor, or enterprise buyer will ask what a given request sent, or which control stopped it.
- Cost is monitored after the fact rather than capped at request time.
When you need a control plane
- Prompts may carry personal, customer, or regulated data that must be blocked or redacted before a provider sees it.
- You must prove, per request, which policy was enforced — for an EU AI Act readiness file, a SOC 2 examination, or a security review.
- Spend must be gated before a request runs, not reconciled at month end.
- Changes to prompts, models, or policies should pass quality gates before they reach production.
How Norcaster fits
You do not choose between the two. Norcaster’s ingress is a drop-in gateway: point an OpenAI or Anthropic SDK at Norcaster, send the X-Norcaster-Deployment-Id header, and the deployment’s policy runs before provider inference. Teams that prefer a structured contract use the run API instead; both paths produce the same enforcement records and traces.
| Route | Purpose |
|---|---|
| POST /v1/chat/completions | OpenAI-compatible chat |
| POST /v1/messages | Anthropic-compatible messages |
| POST /v1/embeddings | Embeddings |
| GET /v1/models | Model listing |
| POST /run | Norcaster's structured run API |
Beyond the gateway, the same runtime governs workflows, retrieval, and models you host elsewhere, and the control plane is where deployments, policies, providers, evals, and evidence exports are managed. It runs EU-hosted or self-hosted, and by default retains decision-level records rather than prompt text.
Read next: runtime protection with Boundary, gateway routes and integrations, and how proof works.
Common questions
Can Norcaster govern AI coding assistants like Claude Code?
Yes — the gateway is OpenAI- and Anthropic-compatible, so coding assistants (Claude Code, Codex CLI, IDE extensions) route through the control plane with one API key per developer. Policy is enforced before the provider sees the prompt, blocked and redacted events are recorded per request, and by default only decision-level records are retained — developer prompts are not persisted unless explicitly opted in.
How do auditors verify Norcaster evidence without trusting Norcaster?
Every CSV and PDF audit-bundle export ships with bundle.json, HOW_TO_VERIFY.txt, and a standalone verify_audit_bundle.py script. The script recomputes the manifest content digest and, where trace-chain anchoring is active, confirms segment Merkle roots against the public Sigstore Rekor transparency log — no Norcaster software, account, or API access required.
Does Norcaster make my company EU AI Act compliant?
No platform can. Compliance determinations require qualified legal counsel and, for certain high-risk systems, a conformity assessment by a notified body. Norcaster produces runtime readiness evidence: enforcement records with pinned policy versions (Art. 9), redaction events (Art. 10), hash-chained record-keeping (Art. 12), approval gates and stop controls (Art. 14), and adversarial-input detection (Art. 15). Its status language is deliberate — evidence captured, controls active — never compliant.
Capabilities described as of September 2026 and scoped to supported production paths. Not legal advice, an attestation, or a compliance assessment.