Deployment
Deploy it where your data has to stay.
Run Neural Inverse Govern on your own servers, in your own cloud account, or on a network with no internet access. The controls and the audit record are the same in each.
- Server
- A single Go binary
- Database
- PostgreSQL 15 or later
- Install with
- Docker Compose or Helm on Kubernetes
- Network
- Designed to run with no internet access
Overview
Govern is built to run on infrastructure you control. The server is a single binary with PostgreSQL as its only required dependency, and it is designed to make no calls to external services at runtime. Choose where it runs; what it enforces does not change.
Self-hosted
On your own servers.
For a single host or your own Kubernetes clusters, on premises.
What you runThe Govern server, the web console and PostgreSQL, with Docker Compose on one host or the Helm chart on Kubernetes. Optional planes handle telemetry intake, data-plane traffic and vendor audit polling.
What we runNothing in your environment. We publish releases with their release notes.
Where data livesIn your PostgreSQL database and your object storage, either a shared directory or an S3-compatible bucket. Older audit entries can move to a bucket you own.
How updates arriveYou choose when. Back up, then upgrade: migrations run first, on every Docker Compose start or in a Helm pre-upgrade hook. To go back, restore the backup rather than migrating down.
RequirementsPostgreSQL 15 or later (the Docker Compose stack ships PostgreSQL 17). Docker 24.0 and Compose v2.20 or later with 2 GB of memory and 10 GB of disk, or Kubernetes 1.29 or later with Helm 3. A reverse proxy or Ingress for TLS.
Keys and secretsYou generate the deployment keys and keep them in your secret store. In production the server refuses to start with missing or placeholder values.
In your cloud
In your own cloud account.
On your Kubernetes cluster, next to a managed PostgreSQL service, your object storage and your secret store.
What you runThe Helm chart on your cluster. Each plane runs as its own workload, non-root with a read-only file system, spread across zones, with network policies that deny ingress by default.
What we runNothing in your account. We publish releases and the Helm chart.
Where data livesIn your account: a managed PostgreSQL database and your bucket, encrypted with your KMS key. The bucket is never public; downloads go through the API, which checks access and audits each one.
How updates arriveYou upgrade with Helm to a pinned release; the chart refuses an unpinned tag. Hooks run the migrations first, and pods roll over without taking capacity away.
RequirementsKubernetes 1.29 or later and Helm 3. PostgreSQL 15 or later reachable from the cluster, with TLS required by default. For more than one server replica, a shared storage class such as EFS, Azure Files or Filestore, or an S3-compatible bucket.
SecretsSecrets never go in chart values. Create them as Kubernetes Secrets, or sync them from AWS Secrets Manager with the External Secrets Operator.
Air-gapped
On a network with no internet access.
Everything Govern needs after installation, from container images to identity, mail and model servers, can live inside your boundary.
What you runThe same stack, installed from an offline bundle of images, Compose files and Helm charts with checksums. Your internal registry, identity provider, mail relay, model servers, DNS, time and certificate authority sit inside the boundary.
What we runNothing. We sign offline licences. Your installation holds only the public key to check them and refuses a licence it cannot verify.
Where data livesInside your network. Audit archiving works with an on-premises S3-compatible store that supports Object Lock, and audit verification needs no network access.
How updates arriveA new bundle is built on a connected host, moved through your approved media process and checked against its checksums inside the gap. Images load with pulls turned off.
RequirementsPostgreSQL 15 or later. An internal registry, or images imported on every node. Object storage on a shared directory or an S3-compatible store inside the network.
ModelsSelf-hosted model servers inside networks you list. Restricted work can be limited to them, and workload identity accepts key sets you upload for issuers Govern cannot reach.
Verify before you rely on itRun each release on a host with no route out and record every connection attempt. A strict air gap is enforced by your network controls; these tests are defence in depth.
Hosted by us
Or we run it for you.
The same software and controls, operated by Neural Inverse.
What you runNothing to install. Your people sign in and use it.
SharedYour organisation on our shared service at cloud.neuralinverse.com, in AWS US East (N. Virginia) or AWS Europe (Frankfurt). Your content stays in the region you choose; sign-in details (email, name, credentials, sessions) and billing records are held in the United States.
DedicatedAn installation for your organisation alone, on AWS, Microsoft Azure or Google Cloud, in the region set in your contract, with its own sign-in system so identity data stays in that region too. Use your own domain, keep encryption keys in your own cloud account, and choose a single region if you need one.
What we runThe Govern installation, its database and storage.
Where content livesCode context, prompts, vault secrets, logs and audit content are created and kept in your region and never copied out of it. Every audit entry records the region that wrote it, and tokens from one region are refused by another.
Who controls itWe operate the service and process your data as your processor, under the data processing addendum. Our staff reach your data only through audited emergency access. To keep it entirely in your hands, run it yourself.
How updates arriveWe apply them.
Compare
The options side by side.
| Self-hosted | In your cloud | Air-gapped | Hosted by us | |
|---|---|---|---|---|
| Operated by | You | You | You | Neural Inverse |
| Runs on | Your servers | Your cloud account | Your isolated network | Shared: AWS N. Virginia or Frankfurt. Dedicated: AWS, Azure or Google Cloud, region of your choice |
| Install method | Docker Compose or Helm | Helm | Offline bundle, then Compose or Helm | None |
| Database | Your PostgreSQL 15+ | Your managed PostgreSQL 15+ | Your PostgreSQL 15+ | Operated by us |
| Updates | When you choose | When you choose | By transferred bundle | Applied by us |
| Data location | Your database and storage | Your cloud account | Inside your network | Content in your home region; on the shared service, sign-in details in the United States |
| Who holds your data | Only you | Only you | Only you | Neural Inverse, as your processor |
In every option
Where it runs changes. What it enforces does not.
The same software
Every option runs the same Govern code paths. A self-hosted installation is one region of its own.
Identity and policy
Single sign-on, roles, model access, agent policy and approvals behave the same wherever Govern runs.
Enforced on the server
Model requests, MCP tool calls and agent-to-agent calls are decided in Govern's gateways, and decisions fail closed.
Audit record
Every change and decision goes into the same hash-chained, append-only log, with signed exports you can verify offline.
Tenant isolation
The database keeps organisations apart with row-level security, and the server refuses to start on a role that could bypass it.
Encryption
Secrets are encrypted at rest under a key hierarchy, and policy delivered to machines is signed.
Guides
Install, operate and review.
Deploy and operate →
Requirements, planes, secrets, health checks and upgrades.
Docker Compose →
A single host, from first start to production.
Kubernetes →
Every plane with the Helm chart, on premises or in your cloud.
Air-gapped →
Build, transfer, install and verify a release offline.
Security architecture →
Isolation, encryption, audit and outbound controls.
Security and trust →
How we protect your code and data, and how to report a vulnerability.
Talk to our team.
Neural Inverse is available by request. We work with your engineering, security and procurement teams from the first conversation.