Run AI agents on your own Kubernetes: models, agents and tools in one Helm install, with every call made as the person who asked.
The Agent Platform takes proven open-source projects, such as kagent, llm-d, agentgateway and Flux, and packages them as a single platform for running AI agents in production. One OIDC identity spans the whole stack, so agents call tools and models as the person who invoked them, never as a broad service account.
It is the agent platform that runs on every Giant Swarm management cluster. The same chart installs on kind, EKS, AKS, GKE or any conformant cluster.
Most components can be switched on or off on their own. A default install runs Tool Access (muster, agentgateway and Valkey) plus model-manager. You add agents, model serving and the portal when you need them.
Declarative agents that each run in their own sandbox, are suspended to object storage when idle and resume on the next request.
| Component | What it does |
|---|---|
| kagent | Kubernetes-native agent framework: agents, templates and model configs as CRDs, plus the controller that runs them. The platform runs Giant Swarm's kagent line. |
| Agent Substrate | The actor runtime under kagent. Each agent is a gVisor sandbox in a worker pod, with its state snapshotted to object storage. |
| agent-manager | Creates, updates and deletes agents as GitOps releases of the agent chart, over MCP and REST, as the calling person. |
| klaus-gateway | Slack front door. A Slack thread becomes an agent session, and the agent acts as the Slack user. |
| workspace-manager | Workspaces: an Organization's repositories mirrored on a shared volume, with a working copy per agent session, served as MCP tools. |
| agent-sandbox | The Kubernetes Sandbox API for isolated, stateful single-pod workloads. |
One MCP endpoint for every tool. Each call carries the caller's own token down to the backend, and toolsets narrow what an agent sees without ever widening it.
| Component | What it does |
|---|---|
| muster | MCP gateway that aggregates all MCP servers behind one endpoint, with OAuth sign-in per server, token forwarding, toolsets and workflows. |
| agentgateway | Data plane for MCP, A2A, gRPC and LLM traffic, with JWT validation, routing policies and telemetry at the edge. |
| mcp-kubernetes | The Kubernetes API as MCP tools, called with the signed-in user's own RBAC. |
| agent-platform-mcps | Registers the platform's MCP servers with muster and agentgateway from one list. |
| Valkey | Session and token store for muster's OAuth server. |
Serve open-weight models on your own GPUs, or use hosted providers such as Anthropic, OpenAI and Gemini. Either way, models are wired into agents for you.
| Component | What it does |
|---|---|
| llm-d | Distributed inference on vLLM with cache-aware routing. The platform ships curated, benchmarked serving presets for it. |
| KServe | The LLMInferenceService API and controller that run llm-d deployments. |
| NVIDIA GPU Operator | Device plugin, GPU feature discovery and DCGM monitoring on GPU nodes. |
| model-manager | Model inventory, pull, load and unload across KServe, Ollama, LM Studio and Lemonade, plus the agents' model configs. |
| cluster-manager | Creates GPU node pools and switches model serving on per cluster, as the calling person. Giant Swarm clusters only. |
| Component | What it does |
|---|---|
| Backstage | The portal for people: browse tools, create and chat with agents, manage models. |
| Flux | Delivers every component and keeps it current. The chart brings its own Flux Operator where a cluster has none. |
| CloudNativePG | Postgres for agent sessions and the Substrate control plane, with backups. |
| vm-manager | KVM virtual machines with a vTPM and attestation, created over MCP. |
| Dex | One identity for people, MCP clients and agents. Other OIDC providers are not supported yet. |
- Identity end to end. One OIDC provider signs people in to the portal, MCP clients and agents. The caller's token reaches the Kubernetes API and every other backend, so RBAC you already have applies to agents too.
- Sandboxed by default. Every agent runs as an isolated actor in a gVisor sandbox, with egress through a controlled gateway and network policies on every hop. The threat model is in Agent Substrate security.
- One install, then GitOps.
helm installbrings its own Flux where the cluster has none, and from then on the release manages itself. Components follow version ranges and roll forward without a new chart release. You can also pin them to a bill of materials. - Adapts to the cluster. The chart detects Cilium, Kyverno and Prometheus and renders for what the cluster actually serves, so a laptop and a production fleet install the same chart.
- Spans clusters. A GPU cluster can run only the serving slice, and a workload cluster only the agent runtime slice, both managed from one installation.
On your laptop: AgentLab runs the whole platform on kind with one binary, including the portal, agents, models and a bundled identity provider:
go install github.com/giantswarm/agentlab@latest
agentlab upOn a cluster: you need Kubernetes ≥ 1.35, Helm 4, the Gateway API CRDs and an OIDC provider. Start from the example values for your cluster shape:
kubectl apply -f https://github.com/kubernetes-sigs/gateway-api/releases/download/v1.5.0/standard-install.yaml
helm install agent-platform oci://gsoci.azurecr.io/charts/giantswarm/agent-platform \
--namespace agent-platform --create-namespace -f values.yaml --wait --timeout 10mThen point any MCP client at https://muster.<your-domain>/mcp. Agents also need the Substrate feature gates, a database and an object store. The installation guide covers each cluster shape step by step.
| Page | What it covers |
|---|---|
| Installation guide | Cluster shapes, prerequisites, identity provider setup, first install |
| Operator reference | Every component and value: GitOps engine, self-management, serving and runtime slices, network policies, observability, CRD lifecycle |
| Authentication | The request path, OAuth discovery, token forwarding and exchange, edge JWT validation |
| Toolset presets | Choosing which tools an agent gets |
| Serving presets | Adding and benchmarking a model recipe |
| ChatGPT login | Running codex on a ChatGPT subscription |
| Agent Substrate security | What the agent runtime changes on the cluster, and what protects it |
| Upgrading | Version-to-version upgrade notes |
Each component is developed in its own repository, linked above. This repository holds the integration that turns them into one platform, and it welcomes contributions too:
helm/agent-platformis the chart you install. It holds the component roster with version ranges, the bundled Flux engine, self-management, and the install and upgrade hooks.helm/agent-platform-connectivitywires the components together: routes and gateways, network policies for Cilium and plain Kubernetes, Kyverno exceptions, the kagent model catalog, serving presets and Grafana dashboards.testsholds the render assertions behindmake verify-*and an end-to-end install on kind (tests/ats).
make verify-all # the render assertions CI runs: Helm 3.17.3, no cluster neededIssues and pull requests are welcome. Please sign off your commits under the DCO.
Please report vulnerabilities through Giant Swarm's responsible disclosure. See SECURITY.md.