Sovereign AI • Open Source
Sovereignty doesn’t stop at the model
You can now run frontier AI entirely inside your own environment, even one that never touches the public internet. Keep the agents, tools and model traffic built on it inside that boundary too, with open source kagent and agentgateway.
Sovereign agentic AI
Sovereign agentic AI rests on four pillars
Keeping data in-country isn’t enough. A sovereign AI system has to answer for where the data lives, who runs the platform and who controls the software. Once agents act on your behalf, it also has to answer for what they’re allowed to do.
Where data lives
Data sovereignty
Confidential data, and the models that reason over it, stay inside your environment. Google Distributed Cloud runs Gemini on premises or fully air-gapped.
Who runs it
Operational sovereignty
You run the whole agentic system: agents, gateway, policy and audit. Everything runs in-cluster, with no external control plane and no required outbound connection.
Who controls the code
Software sovereignty
The code is auditable and community-governed. agentgateway is hosted by the Agentic AI Foundation under the Linux Foundation, and kagent is a CNCF project.
What agents may do
Agentic sovereignty
Sovereignty must follow agency. You control the authority you delegate to agents: what they may access, decide and do, under which policy, and with what evidence. kagent gives every agent a verifiable identity, and agentgateway enforces policy and logs every call.
The agent layer
Every sovereignty question comes back to the agent layer
Agents reason over confidential data, call tools, talk to other agents and act on critical systems. In a disconnected environment, every one of those needs an answer that doesn’t rely on an outside service.
Problem
“Our models are private, but our agents still depend on outside services.”
kagent and agentgateway run entirely in-cluster, with no SaaS control plane and no required outbound connection. That’s the difference between software that runs on premises and software that runs air-gapped.
Problem
“We can’t see or control what our agents send out.”
agentgateway is one point of ingress and egress for agent-to-model, agent-to-tool (MCP) and agent-to-agent (A2A) traffic, with identity, authorization, guardrails, rate limits and token budgets. Every call is logged.
Problem
“We don’t want our agent layer tied to one vendor’s control plane.”
Both projects are open source under neutral governance, so the code can be reviewed and certified. The stack stays as portable and vendor-neutral as the infrastructure underneath it.
Validated on Google Distributed Cloud
Built for disconnected environments
Solo.io and Google Cloud validated kagent and agentgateway on Google Distributed Cloud. Both are available through the Google Cloud Ready – Distributed Cloud initiative.
GDC air-gapped
kagent and agentgateway deploy to standard clusters and connect to Gemini models running entirely on premises.
GDC connected API
Deploy to a Kubernetes cluster alongside the GDC environment hosting Gemini, and use Gemini through its local endpoint.
No SaaS control plane
Identity, policy, telemetry and audit all stay in-cluster. Nothing has to call home to work.
Runs without Kubernetes
Solo Enterprise for agentgateway runs standalone on VMs, bare metal and edge systems with the same policy and MCP controls.
Self-hosted models
Route to models you run yourself, such as Ollama and vLLM, through one OpenAI-compatible endpoint, alongside Gemini on GDC.
Neutral governance
kagent is a CNCF project. agentgateway is hosted by the Agentic AI Foundation under the Linux Foundation, with 300+ contributors across 60+ organizations.
How to do it
Five steps to govern AI inside your boundary
Moving the model in-house is only the start. Governing the agents, tools and traffic built on it is what keeps data in and compute costs down.
Step 1
Pick your environment mode
A connected setup is easier to patch but needs governed egress. A fully air-gapped setup isolates best but makes upgrades harder. Plan for controlled egress rather than none.
Step 2
Right-size your models
Hardware is the hard limit. Run frontier models locally where you can, and pair them with open-weight and task-tuned models sized to your GPUs and memory.
Step 3
Run agents as managed workloads
Give agents declarative deployment, lifecycle management, tracing and evaluation. On Kubernetes, kagent adds the agent layer Kubernetes lacks.
Step 4
Enforce everything at one gateway
Send every model, tool and agent call through agentgateway, with workload identity, egress authorization, MCP and A2A payload inspection, and full logging.
Step 5
Control cost and density
Meter every token for chargeback and per-agent budgets. Suspend idle agents with Agent Substrate so reserved hardware isn’t wasted.
Try it
Start on a laptop
Everything in the stack is open source. Route your existing coding assistant through agentgateway to see token usage before you move to production.
FAQ
Sovereign AI, answered
What is sovereign AI?
Sovereign AI means an organization keeps control of its AI systems: the data, the models, the infrastructure they run on, and the software that operates them. For agentic AI, that control has to extend to the authority delegated to agents: what they may access, decide and do, under which policy, and with what evidence.
What are the four pillars of sovereign agentic AI?
Data sovereignty (where data and models live), operational sovereignty (who runs the platform, with no external control plane), software sovereignty (open, auditable software under neutral governance), and agentic sovereignty (control over the authority delegated to agents: what they may access, decide and do, under which policy, and with what evidence).
Can kagent and agentgateway run fully air-gapped?
Yes. Both run in-cluster with no SaaS control plane dependency and need no external connectivity. On Google Distributed Cloud air-gapped, they deploy to standard clusters and connect to Gemini models running entirely on premises.
Do I need Kubernetes?
Not for the gateway. Solo Enterprise for agentgateway runs standalone on VMs, bare metal and edge systems. kagent and Agent Substrate are Kubernetes-native, and most private and air-gapped platforms, including Google Distributed Cloud, already run Kubernetes.
Which models can I use?
On Google Distributed Cloud, Gemini models alongside local open models. Outside it, agentgateway can route to self-hosted models such as Ollama and vLLM as well as hosted providers, through one OpenAI-compatible endpoint.
Why does open source matter for sovereignty?
Open source under neutral governance lets you audit the code, avoid depending on a single vendor's agent control plane, and keep the stack portable. agentgateway is hosted by the Agentic AI Foundation under the Linux Foundation, and kagent is a CNCF project.
How do I get deployment guides for my environment?
Contact the Solo.io team or your Google Cloud representative. Guides cover both the air-gapped and connected API configurations of Google Distributed Cloud.
Go deeper on sovereign AI
The webinar, the joint blog with Google Cloud, and how to run the gateway anywhere.
Bring your agents inside the boundary
Talk to Solo.io about deploying kagent and agentgateway in sovereign, regulated, or disconnected environments.
