Enterprise AI Gateway & Control Plane

The control plane for every AI request.

Connect private and public AI models behind one governed workspace and one OpenAI-compatible API. Tensor Cortex decides which models each user, application, and data class can use before optimizing for cost or speed.

Request a pilot Explore the platform

Customer-hosted
Single-tenant
Bring your own model keys
Metadata-only audit by default

Route decisions simulation

policy v18 fail-closed content logging off

Incoming

req_tc_8d4a Quarterly variance analysis Finance workspace · Confidential

req_tc_9e12 Summarize this incident On-call service · Internal

req_tc_77c1 Draft the board memo CFO workspace · Confidential

Policy gate

Evaluate finance-private-only standard-routing private-only · primary unhealthy

Deployments

customer vpc corp-glm-primary selected

approved cloud openai-enterprise selected

public fallback outside trust boundary blocked

req_tc_8d4a → corp-glm-primary selected · private · 31 ms

req_tc_9e12 → openai-enterprise selected · approved-cloud · 48 ms

req_tc_77c1 → safe error returned public fallback blocked · prompt stayed private

01 / Platform

Private and public AI, governed as one system.

Tensor Cortex turns identity, application context, data sensitivity, model capabilities, health, and budget into one explainable route decision. Security filters the candidate set first. Optimization only happens inside that set.

01AuthenticateUser or service account
02ClassifyPublic to Restricted
03EnforceDeny and trust boundaries
04FilterCapability and health
05SelectPriority, cost, or latency
06ExplainRoute and audit metadata

Workspace

One AI workspace. No provider decisions for employees.

Employees use company-published profiles such as Automatic, Private, Fast, Coding, or Finance. Tensor Cortex applies company policy in the background and shows a clear security classification on every response.

OIDC single sign-onTeams and roles follow corporate identity.
Privacy can only move upUsers may request Private mode, never bypass policy.
History follows policyConversation storage can be disabled or encrypted with retention.

Compare the Q2 variance notes with our operating plan.

PrivateAutomatic · policy managed

The main variance is concentrated in infrastructure and contractor spend. I found three material drivers...

Ask a follow-up Confidential

Gateway

One stable API while models and providers change.

Applications connect to logical profiles, not physical model IDs. Move the code or automatic profile to a different deployment without shipping application changes.

OpenAI-compatibleChange the base URL and API key.
Streaming and tool callsNormalized across supported providers.
Scoped service accountsApplication identity, profile access, and trusted metadata.

PythonRequestRoute headers

from openai import OpenAI

client = OpenAI(
    base_url="https://ai.company.com/v1",
    api_key="tc_live_..."
)

response = client.chat.completions.create(
    model="automatic",
    messages=[{
        "role": "user",
        "content": "Summarize this incident."
    }]
)

X-TC-Route-Classapproved-cloudX-TC-Fallbackfalse

Control

Model access, routing, audit, and spend in one control plane.

Publish policy without changing application code. Test a user, team, profile, and data class before release, then inspect every selected and rejected deployment through the route decision record.

Versioned policyDraft, simulate, publish, and roll back.
Deployment healthActive and passive signals with circuit protection.
Usage and budgetsAttribute tokens and estimated cost to teams and apps.

WHEN

Teamis Finance

Data classis Confidential

THEN

Trust zonecustomer-vpc or on-prem

Content loggingdisabled

SIMULATION

Profile: Automatic
Result: Allowed
Selected: corp-glm-primary
Rejected: 2 deployments

Same evaluator as production routing

02 / Security invariant

A model failure never becomes a data boundary failure.

Fallback is evaluated against the same policy-approved deployment set. Once a stream starts, Tensor Cortex never silently continues the answer from another provider.

Review the security architecture

Confidential requestPrivate-only policy

corp-glm-primaryconnection timeout

failed

corp-glm-backupcircuit open

unavailable

Blocked by policy Public fallback denied The prompt was not sent outside the approved trust boundary. The application received a safe, explainable error.

Built for governed use

Enterprise controls without a platform team-sized build.

IdentityOIDC, teams, RBAC, service accounts

Model catalogProviders, deployments, trust zones, capabilities

ProfilesAutomatic, Private, Fast, Coding, Finance

PolicyUsers, teams, apps, data classes, regions, tags

RoutingFixed, priority, least cost, lowest latency

ResilienceHealth checks, timeouts, retry, circuit breaker

AuditDecision reasons, rejected candidates, fallback history

EconomicsTokens, estimated cost, team and application budgets

03 / Deployment

Your AI control plane runs in your trust boundary.

Each customer receives a dedicated, single-tenant Tensor Cortex deployment. Provider credentials, route metadata, and optional conversation history stay inside that environment.

Enterprise access Employees · Applications OIDC · Service accounts

Customer environment Tensor Cortex Workspace · Gateway · Control

PrivateVPC / on-prem models

Approved cloudPublic model providers

Customer KubernetesRun in your cluster with Helm, private networking, and your observability stack.

Customer cloud or on-premKeep the gateway and prompt path inside your VPC or data center.

Dedicated managed instanceA separate environment operated for your organization, never a shared data plane.

Focused pilotStart with Docker Compose, one public provider, and one private endpoint.

Policybeforeoptimization

Cost and latency only rank models already allowed by policy.

Profilesbeforeproviders

Users and applications see stable capabilities, not model churn.

Metadatabeforecontent

Audit is useful without collecting every prompt and response.

Explanationbeforeautomation

Every route records what was selected, rejected, and why.

04 / FAQ

Enterprise AI gateway questions, answered.

Technical and deployment questions teams ask before standardizing AI access.

What is Tensor Cortex?

Tensor Cortex is an enterprise AI access and control plane. It gives employees one governed AI workspace, applications one OpenAI-compatible gateway, and IT and security teams one place to manage model access, routing policy, safe fallback, audit, usage, and budgets.

How is Tensor Cortex different from a basic LLM gateway?

A basic LLM gateway proxies requests or selects a provider. Tensor Cortex evaluates who is making the request, which application it belongs to, the data class involved, and the active company policy before it considers cost or latency. The same policy system governs both the employee workspace and developer API.

Can it route between private and public AI models?

Yes. Tensor Cortex connects private OpenAI-compatible endpoints, models in a customer VPC or on-prem environment, enterprise cloud endpoints, and approved public providers. Logical profiles such as Automatic, Private, Fast, or Coding hide physical model changes from users and applications.

What happens when a private model is unavailable?

Fallback stays inside the policy-approved trust boundary. If a confidential request is private-only and no compliant private deployment is available, Tensor Cortex returns a safe error and records the route decision. It does not silently send the prompt to a public provider.

Is the gateway compatible with the OpenAI API?

Tensor Cortex exposes an OpenAI-compatible chat completions endpoint with streaming and logical model aliases. Existing applications can normally connect by changing the base URL and API key while keeping a stable model name such as automatic or code.

Where is Tensor Cortex deployed?

The first production model is single-tenant. Tensor Cortex can run in your Kubernetes environment, cloud account, VPC, or on-prem network. A dedicated managed instance is also available. Docker Compose supports focused pilots.

Does Tensor Cortex store prompts and responses?

Audit is metadata-only by default. Prompt and response content is not written to audit logs unless an organization explicitly enables encrypted content retention. Workspace conversation history can also be disabled by policy.

Does Tensor Cortex deploy or train models?

Tensor Cortex is the control layer, not an inference engine or GPU scheduler. It connects to model endpoints you already operate or procure. Private model deployment and managed inference can be delivered separately from the core platform.

Tensor Cortex Insights

Responsibility-first research for AI access infrastructure.

Evidence-led technical guides that define decision boundaries, cite current primary sources, and include original frameworks or testable artifacts.

View all Insights

Platform and buying architecture Updated July 17, 2026

LLM Gateway, Proxy, Router, and Control Plane: Where Each Responsibility Begins and Ends

A responsibility matrix for separating HTTP forwarding, model selection, policy, configuration, audit, and operations in an LLM access stack.

Read the article

05 / Pilot

Route your first governed request.

Tell us about your environment and we will scope a focused pilot: your identity provider, one public provider, one private endpoint, and your data policies — inside your trust boundary.

Starts smallDocker Compose, one public provider, one private endpoint.
Security-first walkthroughPolicy model, audit records, and fail-closed behavior on real requests.
No platform rebuildApplications keep the OpenAI API surface they already use.

Prefer email? hello@tensorcortex.com