On-Premise · Self-hosted Agent Gateway

Your gateway.
Your infrastructure.
Your data.

Edgee's AI Gateway now runs as a fully on-premise deployment. The same advanced compress, route, and observe capabilities that power our cloud — running entirely inside your own environment. Self-hosting is a deployment choice, not a trade-off against the product.

BYOK · Docker Compose or Helm · connected or fully air-gapped

  • 1

    Container image

    Non-root, read-only root filesystem. Runs anywhere you run containers.

  • 15s

    Config sync

    Connected mode pulls model registry, routing, and spend limits from Edgee.

  • 0

    Prompts sent to Edgee

    Prompts, completions, and provider keys never leave your infrastructure.

  • 1.24+

    Kubernetes

    Helm chart for any modern cluster. Or Docker Compose on a single VM.

When inference has to stay inside your walls

When you process highly sensitive data — financial transactions, clinical records, government case files — maintaining full control over data residency is a core security requirement. All inference must occur within trusted infrastructure. On-Premise lets you connect to the latest AI models and manage traffic across providers without ever relinquishing control over your data, provider keys, or observability.

What you get

One image. Two ways to run it. Nothing stripped down for the on-prem path.

Docker Compose or Helm

A Compose file for a single VM, or a Helm chart for any Kubernetes 1.24+ cluster. Installs like anything else in your stack.

One hardened image

A single container image, running non-root with a read-only root filesystem. Versioned releases, no surprise upgrades.

Nothing stripped down

The same token compression, provider routing, and request-level observability as the hosted gateway. The on-prem path is not a lite edition.

Per-org deployment secrets

A license and signature key, issued per-org in Console → Org settings → On-Premise by an org admin. Pull them from a Kubernetes secret — nothing sensitive in a values file.

Connected or fully air-gapped

The default is connected mode: centralised config, without leaving your network. For environments where no outbound calls to Edgee are allowed, headless mode disables all syncing. Either way, prompts and provider keys stay on infrastructure you control.

Connected

Default

Centralised config, without leaving your network.

  • Syncs model registry, routing rules, and spend limits from Edgee every 15 seconds
  • Reports usage back to keep your team dashboard live
  • Outbound only to your LLM provider and Edgee’s control plane
  • Prompts, completions, and provider keys never leave your infrastructure

Headless

Air-gapped

No outbound calls to Edgee, at all.

  • All syncing disabled — configuration is managed locally from a file
  • No usage data is ever sent to Edgee
  • Route usage logs to your own internal OTLP endpoint for observability
  • Communicates only within your network

What actually leaves your network

In connected mode, the only outbound traffic is to your LLM provider (for inference) and to Edgee's control plane (for configuration sync and telemetry such as usage metrics and health checks).

No prompts, completions, or provider API keys are ever sent to Edgee. Your admins keep the same console they already use — just pointed at a gateway running on your infrastructure.

Control stays where it already is

On-Premise is BYOK end-to-end. The gateway runs your rules, on your keys, against your backends.

  • Your own provider API keys (BYOK) — Edgee never holds them.
  • Your own model allow-list and per-team or per-user spend caps, authored locally.
  • Your own observability backend — route usage to your internal OTLP endpoint instead of the Edgee dashboard.
  • Your own secrets management — the Helm chart pulls license and signature keys from an existing Kubernetes secret, so nothing sits in version control.

What this means for your team

Compliance & security leads

A straight answer to "where does our data go": nowhere it isn't already allowed to go. Prompts and provider keys stay on infrastructure you control.

Platform & infra owners

A gateway that installs like anything else in the stack — helm upgrade --install or docker compose up, versioned releases, no surprise upgrades.

Engineering leaders

The same team dashboard your admins already use, just pointed at a gateway running on your own infrastructure. Fleet visibility without the data leaving your network.

A distinct tier — pick where it runs

On-Premise sits alongside Edgee's Shared Cloud and Dedicated Tenant hosting options. Same product, different operating boundary.

Shared Cloud

Managed, multi-tenant.

  • Fastest path to production
  • Runs on Edgee’s edge
  • Zero infrastructure to manage

Dedicated Tenant

Managed, isolated.

  • Your own isolated tenant
  • Custom data residency (EU or US)
  • Still operated by Edgee

On-Premise

This page

Self-hosted, your infrastructure.

  • Runs entirely inside your environment
  • Connected or fully air-gapped headless mode
  • Prompts and keys never leave your network

Installs like anything else in your stack

One container image, running non-root with a read-only root filesystem. Deployment secrets are issued per-org in Console → Org settings → On-Premise, by an org admin.

# compose.yml — single VM, connected mode
services:
  gateway:
    image: ghcr.io/edgee-ai/gateway:latest
    ports:
      - "8080:8080"
    environment:
      ENV: prod
      # Issued per-org in Console → Org settings → On-Premise
      LICENSE_KEY: ${LICENSE_KEY}
      EDGEE_SIGNATURE_KEY: ${EDGEE_SIGNATURE_KEY}
      TELEMETRY_ENABLED: "true"
    restart: unless-stopped

# docker compose up -d

Full walkthrough for both Docker Compose and Helm in the self-hosting docs.

Frequently asked questions

Bring the gateway inside your walls

If you're evaluating on-premises, the self-hosting docs walk through Docker Compose and Helm end-to-end. If your org needs the On-Premise entitlement enabled, talk to us and we'll get it set up.

Prompts and provider keys never leave your infrastructure.

Edgee Turbo Models - Use Claude Code with Kimi K2.7, MiniMax M2.7, and more | Product Hunt