Munk AI logoMunk AI
Install
Back to Docs
Concepts

Core Concepts

Learn the core concepts behind Munk AI and how its main workflows fit together.

To use Munk AI effectively, you don't need to understand complex testing frameworks or internal architecture. You only need to grasp a few core product concepts that represent what you test, how you structure execution, and what results you get.

1. AppTarget (The Application)

An AppTarget tells Munk AI what it is going to test. It abstracts away the differences between platforms so the AI can interact with them uniformly.

  • For Web: It's typically a set of URLs (e.g., base_url, origin).
  • For Android: It's an application package name (e.g., com.example.app).

You define the target once, and Munk AI figures out how to launch it, capture its UI tree, and interact with its elements.

2. TestCase and RequirementPlan (The Instructions)

Munk AI's formal execution workflow is built around structured testing assets.

The TestCase (Structured Execution)

A TestCase is a structured, reproducible definition of a test. You use it when you have a specific intent and clear success criteria, such as regression testing, recorded flows, or CI pipelines.

  • Components:
    • intent: What behavior is being validated.
    • expected: A list of conditions that must be true for the test to pass.
    • runner_goal: The executable task Munk AI should perform.
  • Used in: The munk run case and munk verify change commands.

The RequirementPlan (Organized Coverage)

A RequirementPlan groups multiple TestCase items into a reviewable, executable asset.

  • Used in: The munk plan, munk run plan, and munk verify change workflows.
  • Why it matters: It is the unit you can review, approve, execute, and track over time.

3. Artifacts & Runs (The Results)

Munk AI doesn't just print "Pass" or "Fail" in the terminal. Because it operates visually and semantically, every execution (a Run) produces structured Artifacts saved in your local workspace (usually under .munk/runs/).

Instead of digging through raw logs, you get:

  • report.json: The final verification verdict (passed, failed, or inconclusive).
  • Screenshots: Visual proof of the app state before and after actions.
  • diagnostics.json: Execution details and any errors encountered during the run.

4. Config & Models (The Brains)

Munk AI is powered by large language models (LLMs), ideally local ones like Gemma 4 to keep costs low and data private.

The Config (typically a yaml file) tells Munk AI which model provider to use (e.g., LMStudio, direct API, OpenRouter) and how to connect to it. You pass this configuration to almost every command using the --config flag, ensuring the "brain" is properly attached to the "hands" (the device runtime).

Next Steps

Now that you understand the basic building blocks, you can dive into specific workflows: