Orchestrate AI agents on Kubernetes. Define your agent(s) (model, prompt, tools) and the operator keeps it running, scaled, and within budget. Chain agents into pipelines with ArkFlow. Trigger them on a schedule or webhook with ArkEvent.
| Resource | What it does |
|---|---|
ArkAgent |
Pool of agent replicas backed by a model, system prompt, and optional MCP tool servers |
ArkService |
Routes tasks to available agents (round-robin, least-busy, random) |
ArkSettings |
Reusable config shared across agents: temperature, output format, prompt fragments |
ArkFlow |
DAG of agent steps where outputs feed into inputs. Supports conditionals, loops, and timeouts |
ArkEvent |
Fires flows on a cron schedule or HTTP webhook, with fan-out to multiple flows |
ArkMemory |
Attaches a memory backend (in-context, Redis, or vector store) to an agent |
# Install the operator
kubectl apply -f https://github.com/arkonis-dev/ark-operator/releases/latest/download/install.yaml
# Deploy Redis (task queue)
kubectl apply -f https://raw.githubusercontent.com/arkonis-dev/ark-operator/main/config/prereqs/redis.yaml
# Add your API key
kubectl create secret generic arkonis-api-keys \
--from-literal=ANTHROPIC_API_KEY=sk-ant-... \
--from-literal=TASK_QUEUE_URL=redis.agent-infra.svc.cluster.local:6379No cluster required. Same YAML, same output.
go install github.com/arkonis-dev/ark-operator/cmd/ark@latest
ark init my-agent
cd my-agent
ark run quickstart.yaml --provider mock --watch
ark run quickstart.yaml --provider anthropic --watch
ark validate quickstart.yamlProvider is auto-detected from the model name: claude-* → Anthropic, gpt-*/o* → OpenAI.
Pre-built binaries available on the releases page.
Full documentation: arkonis.dev
Apache 2.0 — see LICENSE