Plugs into the stack you already run
Karate Agent explores your APIs and web apps and emits ordinary, deterministic, version-controlled Karate tests: reviewable in a pull request, runnable in CI with no agent and no model in the build. One self-hosted container, your LLM, and a safe-to-ship verdict computed by code.
~30s
Scripted flows
8–19 min
Pure LLM approach
72x
Fewer page scans
100% Self-Hosted · Bring Your Own LLM · Docker-Native
Feature: policy rating
Scenario: young driver, sport class
Given path 'quote'
And request { driverAge: 19, vehicleClass: 'sport' }
When method post
Then status 201
And match response.premium == 2417.50
# ordinary Karate: reviewed in a pull request,
# runs in CI with no agent, no model, no tokensPlugs into the stack you already run
The browser is one capability of a capability-agnostic harness; API-only runs never start one. Everything below ships inside the Agent license or alongside it in the Karate platform.
API Testing
REST, gRPC, GraphQL & WebSocket
AI Browser Automation
Display-text locators that survive redesigns
Stateful Mocks
Real logic, never canned payloads
Contract Testing
Does your mock tell the truth?
API Mutation Testing
Would your tests notice a bug?
Test Data Generation
Every row carries the expected result
Business Rules Testing
The rate book, executable and graded
Coverage & Traceability
Exercised vs merely claimed
API Governance
Is the spec fit to ship?
One self-contained image: harness loop · serve console · MCP server · API engine · colocated Chrome. A single karate_eval tool drives it all from Claude Code, VS Code Copilot, or Cursor.
The Mandate
You’ve been told to adopt AI in testing. Karate Agent is how you prove it’s working, safe, and paying back, against the three fears that stall every AI rollout.
“It won’t pay back”
Scripted flows run for zero tokens. Stable explorations converge into deterministic checks that replay for ~$0 in CI, so the meter stops when the authoring stops.
“We can’t trust it”
Coverage is graded exercised versus merely claimed, a violated business-rule guarantee blocks the ship verdict, and every run leaves evidence a reviewer can open. A heal is proposed for a human to approve, never silent. The AI explains the verdict; it never decides it.
“It can’t leave our network”
One container inside your perimeter. Bring your own LLM, including Azure OpenAI inside your existing agreement, or fully air-gapped open-weight models via Ollama.
Single-tenant · RBAC · audit trail with session video · air-gap deployment · zero egress, zero telemetry · license server · one REST call from Jenkins, GitHub Actions, Azure DevOps, or GitLab
The Design Choice
Most AI testing tools make you dependent on the AI to run what it wrote. This one writes itself out of the loop.
Point it at a spec, a base URL, or a screen. The agent drives your API or UI like an engineer would, on your model, over one MCP and REST surface.
What comes out is ordinary Karate: readable, diffable, version-controlled tests your team reviews in a pull request. No proprietary format, no agent lock-in.
Your pipeline runs them with standard Karate: no agent, no model, no tokens, no flake from a model’s mood. The same result, every run.
Tests break when a schema or the DOM shifts. Karate Agent reads the DOM directly instead of screenshotting it, locates elements by visible text so tests survive redesigns, and when a check drifts it re-drives the flow and proposes the repair for a human to approve. A heal never hides a regression, and a deterministic check still runs strict.
Who it serves
The same evidence, read two ways. Product owners audit what the release means; engineers own how it runs.
For Product Owners
For Engineering
Explore Karate Agent
What enterprise teams ask before adopting Karate Agent.
No. The agent emits ordinary Karate: deterministic, version-controlled tests you review in a pull request and run in CI with standard open-source Karate. No agent, no model, and no tokens anywhere in the build. The agent accelerates authoring and maintenance; your pipeline never depends on it.
Both, and more. Karate Agent is not a browser bot; it’s the Karate test runtime in a container. The same self-hosted image runs API testing (REST, gRPC, GraphQL, WebSocket), stateful mocks, business-rule verification, coverage and requirements-traceability, an MCP server, and AI browser automation. The browser is one capability of a capability-agnostic harness. API-only runs never even start a browser.
For many enterprise use cases, yes. Traditional tools like Selenium and Playwright rely on brittle CSS or XPath selectors that break when the UI changes. Karate Agent uses display-text locators and LLM-powered recovery to adapt to UI changes automatically, dramatically reducing test maintenance. Many teams use it as a Selenium alternative or Playwright alternative, especially for complex enterprise SPAs like Guidewire, Salesforce, and ServiceNow.
Claude computer use is a cloud-hosted, vision-based agent that sends screenshots to a vendor and reports “done” on its own say-so. Karate Agent is self-hosted and DOM-based (10–50x more token-efficient), and every run leaves evidence: annotated screenshots, video, rule verdicts, and the real network traffic matched to your API contract. Your data never leaves your network, it integrates with CI/CD via a REST API, and it works with any LLM, not just one vendor.
Karate Agent is LLM-agnostic: bring your own LLM. It works out of the box with Anthropic Claude, OpenAI GPT-4 (including Azure OpenAI), Google Gemini, and any open-weight model served via Ollama, with Gemma 4 26B and Qwen 3.6 35B-A3B benchmark-verified across the full UI-automation suite (Llama, DeepSeek, Mistral, GLM and Kimi also run). You can also connect to any OpenAI-compatible endpoint, including self-hosted vLLM deployments.
Karate Agent is 100% self-hosted. It runs as a Docker container on your infrastructure: no SaaS dependency, no data exfiltration, no telemetry. Paired with a local or Azure OpenAI model, the entire stack runs inside your perimeter. This makes it suitable for regulated industries like financial services, insurance, and healthcare that require full data sovereignty.
Dramatically less than screenshot-based agents. Karate Agent reads the DOM directly rather than sending pixel images to the LLM, 10–50x more token-efficient. Scripted flows consume zero tokens, a look() diffing function reduces page scans by 72x, and stable explorations converge into deterministic checks that replay for ~$0 with no AI in the loop.
Scoped pilots run on your infrastructure, with your approved models, and are measured on the estate you already have: what the agent authors, what it heals, what it covers.
Java 21 and Docker are the only prerequisites.