Distributed Claude AI agents on Kubernetes
A stateless, Kubernetes-native platform for running persistent Claude AI agents. Built on CRDs, operators, and the Kubernetes API — agents are first-class cluster resources.
From pip install to streaming events in 10 lines
Create an agent, stream its tokens, tool calls, and cost — all over one WebSocket. Typed.
# pip install komputer-ai-sdkfrom komputer_ai.client import KomputerClientclient = KomputerClient("http://localhost:8080")# Create an agent and give it a taskclient.create_agent( name="my-agent", instructions="Analyze our K8s cluster and suggest cost optimizations", model="claude-sonnet-4-6",)# Stream events as the agent worksfor event in client.watch_agent("my-agent"): if event.type == "text": print(event.payload.content) elif event.type == "tool_use": print(f" -> using {event.payload.name}") elif event.type == "task_completed": print(f"Done — cost: ${event.payload.cost_usd}") breakKubernetes-native
Agents are first-class CRDs. The operator reconciles them like any other resource — Pods, PVCs, lifecycle, GitOps.
Stateless API
No database. CR `.status` is the source of truth; Redis is just transport. Restart anything, lose nothing.
Real-time streaming
REST for control, WebSocket for live event streams from every agent — token-by-token output, tool calls, costs.
Squads & sub-agents
Multi-agent shared workspaces via co-located Pods. Manager agents create and orchestrate workers via MCP tools.
Built-in observability
Structured logs, Prometheus metrics, ServiceMonitors, and a Grafana dashboard ship with the chart.
MCP connectors
Plug in MCP servers per agent or template. Credentials managed via secrets, lifecycle by the operator.
Watch every agent think, in real time
Stream token-by-token output, every tool call, and per-task cost over a single WebSocket. Same data the SDKs and CLI consume — no polling, no replay loss.
Read the integration guide →