Kubernetes-native
Squads are CRDs. A Go operator reconciles them the way Kubernetes already reconciles Deployments and Jobs — same tooling, same RBAC, same operator surface.
Kubernetes-native · Agent-agnostic · Open source
Your agents, in formation.
KSquad is a Kubernetes-native control plane for running a squad of AI agents against a shared backlog. It treats a crew of agents as a first-class, reconciled workload — the way Kubernetes already treats Deployments and Jobs. You keep your credentials, your source, and your runtime of choice.
helm install ksquad ksquad/ksquad --namespace ksquad-system --create-namespace
What is KSquad
Agent-agnostic
Any coding agent plugs in behind one documented shim contract. KSquad orchestrates the crew; it never tries to be the agent.
Credential-honest
Every agent runs on its own user’s subscription. KSquad never holds a shared master credential, and your source never leaves your cluster.
What makes KSquad different
Orchestrate the agent — don’t reimplement it
Runtime-agnostic. Claude Code, OpenClaw, Hermes and others plug in behind one shim contract (A2A southbound, MCP for tools). Swap or mix runtimes without rewriting your platform.
A reconcile control plane, not heartbeat glue
Every unit of work is a Run — a reconciled workload with a crash-safe state machine. A controller restart never double-drives a Run, and no coordination state is trapped in a dead pod.
Durable, first-class work items
Coordination is a durable Postgres record — checkout, claim, lease, fencing, audit — not agents DMing each other. At most one agent holds a work item at a time; crashed work is safely reclaimed.
Bring your own credentials — and your own model
Each agent authenticates with a per-user Kubernetes Secret on its own subscription. Connect Claude with one OAuth click; a controller keeps the token fresh. Prefer a local model? Point at your own endpoint.
Safe by construction
Untrusted agent code runs in warm-pool sandboxes (gVisor by default) with per-squad namespaces, RBAC, and NetworkPolicy egress control. A compromised agent can’t fabricate a Run or escape its namespace.
Legible from install to incident
A web console shows your fleet, live Run streams over SSE, build artifacts, per-project dashboards, and a full audit trail. Ship telemetry anywhere with an opt-in OTelConfig CRD.
One afternoon to first squad
One helm install brings up the operator, apiserver, memory service, console, Postgres, and event bus. Sane defaults and a quickstart get you from empty cluster to first running squad in an afternoon.
How it works
Install the control plane
helm install KSquad. You get the operator, apiserver, memory service, and console, with Postgres and NATS bundled as boring subcharts.
Connect a credential & define agents
Click Connect Claude once, then declare 2–3 Agents from the bundled runtimes. Each has a Role and Skills.
Point at a repo & form a squad
Create a Project (a repo + a workspace) and group your agents into a Team. That’s your squad.
Start a Run
Kick off a Run from the console or with kubectl apply. Watch progress stream live and inspect the artifacts.
See it running
Get started
Empty cluster to first running squad in under an afternoon.
helm repo add ksquad https://k8squad.io/charts
helm install ksquad ksquad/ksquad --namespace ksquad-system --create-namespace