Karate Agent ships as a standard Docker image with Chrome and the LLM runtime baked in. Drop it into Kubernetes, ECS, or Docker Compose — one container per test session, fully isolated, scaled by the orchestrator you already run.
Why Docker
Isolation. Each test session gets its own Chrome instance, its own filesystem, its own cookie jar. State from one test never leaks into another. Selenium Grid users will recognize the pain of shared-host browser pools — containers remove it entirely.
Reproducibility. The Chrome version that runs on a developer’s laptop is the exact same version that runs in CI. Pinned image tags eliminate the “works on my machine” class of bugs that haunt non-containerized browser testing.
Scale. Kubernetes treats each session as a pod. Need 500 parallel test runs? That’s a deployment scale-up, not a host-procurement project. Pay only for the minutes you use.
Session isolation
Container start, run test, container die. State doesn’t persist between tests because there’s no shared host.
Old way
Container model
Kubernetes
Standard manifests, no proprietary operators. If your platform team supports any Java/Go service today, they support this.
# karate-agent-deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
name: karate-agent
spec:
replicas: 10 # 10 always-ready pods
selector:
matchLabels: { app: karate-agent }
template:
spec:
containers:
- name: agent
image: karatelabs/karate-agent:1.0
ports: [{ containerPort: 8080 }]
resources:
requests: { cpu: "500m", memory: "1Gi" }
limits: { cpu: "2", memory: "2Gi" }
---
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
spec:
scaleTargetRef: { kind: Deployment, name: karate-agent }
minReplicas: 10
maxReplicas: 500 # scale to 500 in peak hours
metrics:
- type: Resource
resource: { name: cpu, target: { averageUtilization: 70 } }Same pattern works on EKS, AKS, GKE, OpenShift, and on-prem Kubernetes. For smaller deployments, ECS Fargate or Azure Container Instances handle it with a different manifest but the same operational model.
CI integration
REST API in, structured report out. Works the same in Jenkins, GitHub Actions, Azure DevOps, GitLab CI, CircleCI, Tekton, or whatever else you run.
# In any CI tool that can run a shell command
# 1. Submit a test scenario
curl -X POST http://karate-agent.test/sessions \
-H "Content-Type: application/json" \
-d '{
"scenario": "verify checkout flow",
"url": "https://staging.example.com"
}'
# Returns: { "sessionId": "sess-abc123" }
# 2. Wait for completion (or poll)
curl http://karate-agent.test/sessions/sess-abc123/status
# Returns: { "status": "passed", "duration_ms": 4231 }
# 3. Download the HTML report + video
curl -o report.zip \
http://karate-agent.test/sessions/sess-abc123/artifactsThree commands. Works identically across every CI provider. No plugin to install, no proprietary YAML extension. Attach the report.zip to the build artifacts and you’re done.
Headed when you need it
CI · default
Browser runs without rendering to a display. Faster, lighter, uses ~half the RAM, and is what you want for CI. Same DOM semantics, same network stack, same JavaScript engine — just no GUI.
Debugging
When a test fails mysteriously, restart it in headed mode and watch via the built-in noVNC dashboard — live browser session in your web browser. Pause, inspect, inject commands, resume.
Image footprint
~1 GB compressed. Includes Chrome stable, the Karate Agent runtime, JRE 21, and the supporting tooling. Comparable to selenium/standalone-chrome in size.
~30 seconds first pull on a typical CI runner. After that, layer caching reduces start time to under 5 seconds — the time it takes to start Chrome itself dominates.
For air-gapped environments, the image mirrors cleanly to internal registries. No external dependencies fetched at runtime — once pulled, it runs offline. See self-hosted AI testing for the air-gap deployment story.
FAQ
Docker browser testing runs browser automation inside Docker containers — one container per session, isolated from other tests and from the host. It’s the modern foundation for reliable CI/CD test execution: deterministic environments, parallel scale, no version skew between dev and CI.
Three reasons: (1) Isolation — each test session runs in a clean container, no state leaks; (2) Reproducibility — Chrome/Firefox versions pinned via image tags, identical across local and CI; (3) Scale — Kubernetes orchestration runs hundreds of tests in parallel without host-level bottlenecks.
Karate Agent is Docker-native. Each test session runs in its own container with a dedicated Chrome instance — full isolation, zero shared state. Kubernetes manifests and Docker Compose configurations are both supported. A standard REST API triggers sessions, so any CI/CD pipeline integrates with a single curl command.
Yes. Karate Agent ships as a standard Docker image (karatelabs/karate-agent) runnable on any container platform. Kubernetes (EKS, AKS, GKE, OpenShift) is the most common pattern at enterprise scale. ECS Fargate works well for smaller deployments.
Karate Agent runs headless by default (faster, less resource-intensive, ideal for CI). Headed mode is available for debugging — paired with the built-in noVNC dashboard, you can watch the browser session in real time over a web socket, pause it, inject commands, and resume.
Any CI tool with HTTP capability integrates in minutes. Example: curl -X POST http://agent/sessions -d '@test.json', wait for completion, download report. See Enterprise Evaluation for specific pipeline examples across Jenkins, Azure DevOps, GitHub Actions, GitLab CI, and Tekton.
Kubernetes is the standard path. Karate Agent supports horizontal autoscaling — each session runs in its own pod, so adding capacity is adding cluster nodes. Customers running 500+ concurrent test sessions is routine.
Karate Agent is a lean image (~1GB compressed) with Chrome and runtime dependencies. First pull is ~30 seconds on a good connection; subsequent starts use image cache (<5 seconds). Size is comparable to Selenium-Chrome Docker images.
Standard Docker image. Standard Kubernetes manifests. REST API in, JUnit XML out. No platform-engineering escape rooms.