# Deployment models > Where Agent Router runs: fully hosted options (Service and Enterprise Fully Managed) and Enterprise Self-Hosted Data Plane, where you install only the data plane and Tetrate always hosts the management plane. # Deployment models Tetrate Agent Router can be run in three get-started paths. All offer the same gateway and feature set. What changes is where the data plane runs — the part that carries AI traffic — and who operates that infrastructure. Fully hosted paths are the fastest way to start; Self-Hosted Data Plane keeps all AI traffic inside your environment while Tetrate still hosts the management plane. This page explains deployment models from first principles, defining each term as it appears. No prior familiarity with Tetrate Agent Router or cloud infrastructure is assumed. Once the shape of each option is clear, the linked pages at the end go deeper on networking, security, and operations. In every Enterprise deployment, Tetrate hosts and operates the management plane. "Self-hosted" means only the data plane runs in your Kubernetes cluster. You do not install, patch, or scale the management plane — the Console, Admin Dashboard, and configuration backend stay with Tetrate. Deployment model is separate from product tier. Agent Router Service and Agent Router Enterprise describe which features an account has (see the welcome page). This page describes where the software runs and maps to the three get-started paths: Agent Router Service, Agent Router Enterprise Fully Managed, and Agent Router Enterprise Self-Hosted Data Plane. ## The one idea to hold on to first Tetrate Agent Router is built in two halves, called *planes*, that run in different places and do different jobs. - The **management plane** is the control room. It stores configuration (routing rules, policies, provider settings, user accounts) and serves the two web applications operators and developers use. No live AI traffic passes through it. **In every deployment model, the management plane is hosted and operated by Tetrate.** - The **data plane** is the road the traffic travels. Every prompt sent to a model and every response returned passes through it, and nowhere else. The data plane is where deployment models differ. *Prompt* here means the text an application sends to an AI model, and *response* means the text the model returns. Because everything sensitive rides on the data plane (which can include proprietary code, customer records, or internal documents), where the data plane runs is the question that matters most for security and compliance. The split-plane model is covered in full in [Planes and core components](/product-architecture/planes-and-components). The rest of this page is, in effect, the answer to one question: *where should the data plane run?* ## Three paths and who runs each plane | Get-started path | Tier | Management plane | Data plane | Customer installs | | :---- | :---- | :---- | :---- | :---- | | [Agent Router Service](/agent-router-service/quickstarts) | Service | Tetrate | Tetrate | Nothing | | [Agent Router Enterprise Fully Managed](/agent-router-enterprise/fully-managed) | Enterprise | Tetrate | Tetrate | Nothing | | [Agent Router Enterprise Self-Hosted Data Plane](/agent-router-enterprise/self-hosted-data-plane) | Enterprise | **Tetrate** | **Your Kubernetes cluster** | Data plane + gateway only | **Agent Router Enterprise Self-Hosted Data Plane** is a hybrid deployment: Tetrate hosts the management plane; you install only the data plane in your Kubernetes cluster. Service and Fully Managed are both fully hosted — Tetrate runs both planes. They differ by tier (features and scale), not by where software runs. Self-Hosted Data Plane is the only path where you operate infrastructure, and that infrastructure is the data plane alone. ## What you host vs what Tetrate hosts | Component | Fully Managed (Enterprise) | Self-Hosted Data Plane (Enterprise) | | :---- | :---- | :---- | | Management plane (Console, Admin Dashboard, config) | Tetrate | **Tetrate** | | Data plane (gateway, Controller) | Tetrate | **Your Kubernetes** | | Prompt and response traffic | Tetrate infrastructure | **Your environment** | | Customer installs | None | Data plane + gateway only | ## The two Enterprise deployment models at a glance Enterprise offers two deployment models. Both use the same Enterprise feature set; both keep the management plane with Tetrate. Agent Router Enterprise · Fully Managed Management plane: Tetrate · Data plane: Tetrate Tetrate runs both planes. There is nothing for you to install or operate. The gateway is reachable as soon as onboarding completes. Hosted by Tetrate end to end; nothing to provision or install Operational from the moment onboarding completes The model catalog is kept current by Tetrate as providers and models are added Suited to teams that want Enterprise features without running infrastructure Agent Router Enterprise · Self-Hosted Data Plane Management plane: Tetrate · Data plane: your cluster You install the data plane in your Kubernetes cluster. The management plane remains hosted and operated by Tetrate. This suits organisations with data-residency requirements, private network topologies, or a need for deeper control over where AI traffic runs. Data plane deployed in your Kubernetes cluster, in any cloud or on-premises Management plane hosted and operated by Tetrate — no management-plane infrastructure to maintain All prompts and responses stay within your environment and never transit Tetrate systems Full control over model availability, provider credentials, network policy, and data residency Supports private provider connectivity to keep model traffic off the public internet **Self-hosted data plane does not mean:** - You run the Admin Dashboard or Console on your infrastructure - You operate the management-plane backend - Prompts flow through Tetrate-hosted gateway infrastructure **Self-hosted data plane does mean:** - You run the gateway and Controller in your cluster - Tetrate still hosts SSO configuration, the model catalog, users, and audit UI - The data plane pulls configuration outbound from the management plane; Tetrate never connects into your environment ## Agent Router Service (fully hosted) [Agent Router Service](/agent-router-service/quickstarts) is the Service-tier fully hosted option. Tetrate runs the entire platform — both planes. An account is created and the gateway is reachable straight away at a Tetrate-hosted address. There is no cluster to stand up, no software to install, and no infrastructure to keep patched or scaled. Because the data plane is operated by Tetrate, prompts and responses are processed on Tetrate-hosted infrastructure rather than inside your network boundary. For many teams that is the right trade: the gateway works immediately without the infrastructure Self-Hosted Data Plane requires. Service suits individual developers, small teams, evaluations, and workloads without strict requirements about where data is processed. ## Enterprise Fully Managed in depth In the Fully Managed model, Tetrate runs both planes. You configure the management plane through onboarding and operate from the Admin Dashboard or Console. Applications point at the Tetrate-hosted gateway endpoint and send requests using an API shape they already support (Agent Router is OpenAI-compatible, so existing code often works unchanged). Fully Managed suits Enterprise teams that want the full feature set without deploying or operating a data plane in their own cluster. ## Enterprise Self-Hosted Data Plane in depth In the Self-Hosted Data Plane model, only the data plane runs on infrastructure you control. The management plane stays with Tetrate. This is the option when the deciding question is not "does it work" but "where exactly does our traffic go, and who can touch it." A few terms make this model easier to follow: - **Kubernetes** is the industry-standard system for running containerised software (applications packaged so they run the same way anywhere). The data plane is deployed into a Kubernetes *cluster*, meaning a group of machines Kubernetes manages as one. That cluster can sit in any major cloud or in your own data centre. - **On-premises** (often shortened to on-prem) means running software on infrastructure the organisation owns and operates itself, typically in its own data centre, rather than on a cloud provider's infrastructure. Self-Hosted Data Plane supports on-premises deployment of the data plane. - **Data residency** is the requirement that data be processed and stored within a specific geographic region (for example, keeping European traffic inside the European Union). Running the data plane in a location you choose is how a residency requirement is satisfied at the infrastructure level. Residency can also be enforced at the routing level; see [Compliance](/reference/compliance). Because the data plane sits inside your environment, every prompt and every response stays inside that boundary and never transits Tetrate systems. The management plane still supplies the configuration, but it never sees the traffic. The two planes stay in sync over a single connection that the data plane opens *outbound* to the management plane to pull down the latest configuration; the management plane never opens a connection *into* your environment. This outbound-only design means your firewall needs no inbound openings for Agent Router. The connection direction and network posture are covered in [Network and security](/product-architecture/network-and-security). Self-Hosted Data Plane also unlocks capabilities that depend on running inside your network: - **Private connectivity** keeps traffic to AI providers off the public internet by routing it over the cloud provider's own private backbone instead. Each major cloud offers this under its own name: Azure Private Link, AWS PrivateLink, and Google Cloud Private Service Connect (GCP is Google Cloud Platform). When the data plane runs inside one of these clouds, a provider can be configured to be reached over the matching private-connectivity option. See [Network and security](/product-architecture/network-and-security). - **Chargeback** is the practice of attributing AI spend back to the team, project, or cost centre that incurred it, so internal costs can be settled or reported. In Self-Hosted Data Plane there is no payment processing to run; usage is tracked for internal audit and chargeback rather than for billing by Tetrate. - **Guardrails** are inline content rules, for example redacting personally identifiable information or blocking a disallowed request, applied to prompts and responses as they pass through the gateway. Because the guardrails run inside your own data plane, sensitive content is inspected without leaving the environment. Guardrails are covered in [Guardrails](/product-architecture/guardrails). - **Bring Your Own Key (BYOK)** means using the caller's own provider credentials, so that spend is settled directly with the provider rather than through the platform. This works transparently with the routing features in Self-Hosted Data Plane. Self-Hosted Data Plane suits organisations with data-residency or regulatory requirements, teams that require AI traffic on private networks, and deployments that need control over model availability, provider credentials, and network policy. ## Comparing deployment models The management plane is hosted by Tetrate in every model. Differences follow from where the data plane runs. | Dimension | Agent Router Service | Enterprise Fully Managed | Enterprise Self-Hosted Data Plane | | :---- | :---- | :---- | :---- | | Tier | Service | Enterprise | Enterprise | | Management plane operator | Tetrate | Tetrate | **Tetrate** | | Where the data plane runs | Tetrate-hosted infrastructure | Tetrate-hosted infrastructure | Your Kubernetes cluster (any cloud or on-premises) | | Who operates the data plane | Tetrate | Tetrate | You | | Where prompts and responses are processed | On Tetrate-hosted infrastructure | On Tetrate-hosted infrastructure | Inside your environment; never transits Tetrate systems | | Data residency control | Not customer-selected at the infrastructure level | Not customer-selected at the infrastructure level | Data plane runs in a location you choose | | Private connectivity to providers | Not applicable | Not applicable | Supported where the cloud offers it (Azure Private Link, AWS PrivateLink, GCP Private Service Connect) | | Setup effort | None; operational once the account is activated | Onboarding only; no data plane install | Onboarding plus data plane and gateway install in your cluster | | Best suited to | Individual developers, fast starts, evaluations | Enterprise features without running a data plane | Data residency, private networks, infrastructure control | Whichever model is chosen, the gateway, its APIs, and the way applications call it stay the same. Moving from one model to another is a change in where the data plane runs, not a change to the applications built on top of it. ## Which path should I choose? - **Need to start quickly with no infrastructure?** → [Agent Router Service](/agent-router-service/quickstarts) or [Agent Router Enterprise Fully Managed](/agent-router-enterprise/fully-managed) - **Need Enterprise features but don't want to run a cluster?** → [Agent Router Enterprise Fully Managed](/agent-router-enterprise/fully-managed) - **Need prompts and responses to stay in your environment?** → [Agent Router Enterprise Self-Hosted Data Plane](/agent-router-enterprise/self-hosted-data-plane) — management plane still Tetrate-hosted; you install the data plane only Where to go next Agent Router Enterprise Fully Managed Tetrate hosts both planes. Configure onboarding and operate from the Console. Agent Router Enterprise Self-Hosted Data Plane Tetrate hosts the management plane; you install the data plane in your cluster. Network and security Private connectivity, isolation, and transport security for a customer-managed data plane. Operations and reliability How a customer-managed data plane stays available and scales. Planes and core components The split-plane model and the data-plane components that route traffic. Compliance Data-residency options, certifications, and upstream-provider retention settings.