Kubernetes-native
Squads, agents, and runs are CRDs. Deploy with Helm, manage with kubectl. If you can operate a cluster, you can operate a squad — same objects, same RBAC, same reconcile loop.
Multi-agent orchestration for Kubernetes
Your agents, in formation.
Compose squads of AI agents as Kubernetes resources. They claim work, hand off, and coordinate — and you watch every run, live, with OpenTelemetry. Bring your own runtime, your own credentials, and your own model; KSquad reconciles the crew like any other cluster workload.
helm install ksquad ksquad/ksquad -n ksquad-system
What is KSquad?
Agents in formation
Compose teams that claim work, hand off, and coordinate over a shared, durable bus — never peer-to-peer DMs. At most one agent holds a work item at a time, and a crashed agent’s work is safely reclaimed.
Observable by default
Every run emits OpenTelemetry traces and streams live over SSE. No black boxes: follow each tool call, model response, and log line as it happens, and ship the telemetry to any OTLP backend you already run.
Everything you need to run agent squads
01 Project-scoped squads
Every squad lives in its own project and namespace — RBAC-gated and NetworkPolicy-isolated. One team’s agents can’t see, or reach, another team’s work, credentials, or cluster.
02 Agent org views
A live org chart of your crew: who leads, who reports to whom, and what each agent is doing right now. Leadership and role views make a running squad legible at a glance.
03 Build browser
Every artifact an agent produces — diffs, files, logs — is browsable across every Run and addressable by content hash. Inspect exactly what changed, and where it came from.
04 Live runs
Follow a Run as it unfolds over SSE: each tool call, model response, and log line, in order. A controller restart never double-drives a run, so what you see is what actually happened.
05 RBAC
Two global roles and three per-project access levels map cleanly onto Kubernetes RBAC — operators run the platform, authors compose the work, and everyone’s access is auditable.
06 OTel-native
Traces, metrics, and logs out of the box. An opt-in OTelConfig CRD fans each signal out to any OTLP backend — traces to one destination, metrics to another, logs to a third.
07 Plugin SDK
Extend the platform with sandboxed, least-privilege plugins that react to typed events — run.succeeded, workitem.claimed, workitem.handoff, artifact.registered, and more, over durable NATS JetStream subjects. A plugin can never block a run; a flaky endpoint can never stall the platform.
08 Responsive
The full operator console — dashboards, kanban, agent org, build browser — adapts from a wide NOC display down to a phone, so you can check a squad from wherever you are.
How it works
Compose CRDs
Declare a Team and its agents in YAML — each agent with a Role (how it behaves) and Skills (what tools it may use). Point a Project at a repo. kubectl apply.
Squad spins up
The operator reconciles your objects into running agents, scheduling them as pods and wiring the coordination bus. Reconciliation is level-based — edit and re-apply, and the crew converges to match.
Agents work
Agents claim work items, hand off, and coordinate through the durable bus. At most one holds an item at a time; a crashed agent’s work is reclaimed and re-dispatched. The coordination record is the audit trail.
You monitor
Watch runs stream live, follow the traces, and inspect the artifacts — from the console or straight from kubectl. No orchestration code to write.
See it in action
Get started in one command
Install the operator, then open the console.
helm repo add ksquad https://charts.k8squad.io && helm install ksquad ksquad/ksquad