Sovereignty you can enforce in code

Most sovereignty claims are about geography. The servers are in a given country, the data stays in a given region, the contract says so.

That matters, but it is a promise rather than a property. It does not tell you what the software does when someone configures a new AI model, pastes a customer's email address into a chat, or connects a tool that quietly calls an API on another continent.

We wanted the platform itself to answer those questions.

Policy in the request path

AI Hub is the AI layer of OpsPlane: a chat interface for engineers, an OpenAI-compatible API for applications, and agents that use the team's runbooks and platform tools. Every turn passes the same checks, in the same order, before a model is called:

1. Which model may this agent use? Every provider is tagged with where it processes data: self-hosted, EU, or elsewhere. You decide which providers your agents may use. For European teams, the eu_only policy refuses any provider outside the EU or your own infrastructure, with a message saying why. Mistral, the European clouds' model APIs and self-hosted models work out of the box.
2. What is in the message? Email addresses, phone numbers, IBANs (with the checksum verified), card numbers (Luhn-checked) and IP addresses are replaced with placeholders before the text reaches a model or the database. A guardrail can instead refuse such messages outright, or block patterns you define.
3. Can this person afford it? Budgets in tokens or euros, per team, per person or per agent.
4. What may the agent do? Tools that act on the platform, like listing alerts or acknowledging one, run with the calling engineer's own credentials. An agent can never do more than the person using it.

Everything is written to an audit log: every turn, every tool call, every refusal.

Configuration you can review

None of this lives only in a settings page. Providers, agents, guardrails, budgets, knowledge bases and tool servers are Kubernetes resources, like everything else in OpsPlane:

apiVersion: opsplane.io/v1alpha1
kind: AIAgent
metadata:
  name: oncall
spec:
  provider: mistral
  sovereignty: eu_only
  tools: [alerthub_list_alerts, alerthub_get_alert]
  knowledgeBases: [runbooks]

Changing the policy is a pull request, reviewed and reverted like any other change. The operator puts back anything someone edits by hand.

Small enough to run yourself

Sovereignty is easier when you can run the thing. Every OpsPlane service is a single Rust binary; all seven together use under 170 MB of memory at idle. That is the difference between "self-hosting is possible" and "self-hosting is something a small team actually does".

We are looking for a handful of design partners to run it with us. If your team is on call, wants AI without data leaving the providers it trusts, or has to move off Opsgenie before April 2027, we would like to hear from you.