flokoa is the open-source agent harness for Kubernetes
A Kubernetes-native runtime for single agents and swarms of pydantic-ai agents: declarative definitions, packaged capabilities, isolated sessions, event triggers, and an A2A gateway, on your own cluster.
- Single framework, deep integration — built on pydantic-ai primitives, not wrappers around them
- Declarative core — define agents, models, instructions, and tools as Kubernetes resources; edit YAML, behavior changes, no image rebuild
- Boring infrastructure — OCI registries, Argo Events, OpenTelemetry, Helm, lean controllers
Install the operator with Helm (chart published to GHCR on every release):
helm install flokoa oci://ghcr.io/danielnyari/charts/flokoaThen follow the getting started guide.
Boot the entire stack on a local minikube — cluster, operator, gRPC server, Argo Workflows + the A2A executor plugin, a sample agent, and background port-forwards for every UI:
echo "OPENAI_API_KEY=sk-..." > .env # read automatically by `make up`
make upmake up is idempotent and reuses the existing Makefile targets. It will:
-
start (or refresh) minikube and enable the
yakddashboard addon; -
build all images straight into minikube's docker daemon — no registry push, no
image load, and no:latestpull surprises; -
install the CRDs and deploy the operator, server, Argo Workflows and executor plugins;
-
create the
openai-api-keysecret and deploy the petstore sample agent + demo workflow; -
port-forward the three UIs in the background and print their URLs:
UI URL Flokoa UI http://localhost:8080 Argo Workflows https://localhost:2746 (self-signed) yakd (Kubernetes dashboard) http://localhost:8081
Tear it down with:
make down # stop port-forwards + undeploy workloads
make down ARGS=--forwards-only # only stop the port-forwards
make down ARGS=--stop-minikube # also stop the minikube VMKnobs (env vars): WITH_TESTDATA=false skips the sample agent, CONTAINER_TOOL=podman
swaps the builder, and FLOKOA_UI_PORT / ARGO_UI_PORT / YAKD_UI_PORT remap the local
ports. The underlying targets live in operator/Makefile
(make local-up / local-down / local-urls).
- Getting started
- Architecture
- Agent CRD · AgentTool · Model · ModelProvider · AgentTrigger
- Quick reference
- Roadmap — the Pivot v2.1 plan this project is executing
operator/— Kubernetes operator (Go): CRDs, controllers, admission webhooks, gRPC API, Helm chartsdk/python/— Python SDK (uv workspace):flokoaCLI/library, generated CRD types, the generic runner (flokoa-runner)
Early development, executing the Pivot v2.1 roadmap. Phase 0 and P0a — runtime contract, spec compiler, generic runner, virtual endpoint identity, and injected telemetry — are done; the Capability CRD (P0b) is next, and isolated sessions, the A2A session-routing gateway, and swarms (SwarmRun) are future work. See the roadmap for what is frozen, kept, and coming.