MCP Tools, Resources & Prompts: Complete Deep Dive for SDETs and Playwright Engineers

In the previous tutorials we learned MCP Fundamentals (tutorial 1) and the MCP Host, Client & Server architecture (tutorial 2). Now we go deeper into one of the most important parts of MCP: what an MCP server actually exposes to an AI application. The three core concepts you should understand are Tools, Resources and Prompts.

For an SDET or Playwright engineer, this can mean: a Tool clicks a button, a Resource reads a test specification, and a Prompt provides a predefined test-analysis workflow. Understanding this distinction is fundamental to MCP architecture.

1. Where Tools, Resources, and Prompts Fit

Let's start with the full architecture.

The MCP server can expose one or more of these capabilities. The AI application can then discover and use them depending on its needs.

2. What Is an MCP Tool?

An MCP Tool is an executable capability exposed by an MCP server. A tool is used when the AI needs to perform an action.

• Open browser
• Navigate to URL
• Click element
• Type text
• Query database
• Create Jira issue
• Read repository file
• Execute command

3. Playwright Tool Example

Suppose a Playwright MCP server exposes browser-related capabilities. Conceptually, it could expose tools like browser_navigate, browser_click, browser_type and browser_snapshot. The exact names depend on the server implementation.

Now imagine the SDET tells the AI: "Open the login page." The AI reasons: I need browser navigation. Then:

The tool performs the action.

4. Tool Execution Flow

5. Tools Usually Have Inputs

A tool often requires input parameters. Conceptually, the tool browser_navigate takes an input url, and the AI might invoke it with url = https://qa.example.com/login.

Another example — a query_database tool with a query input:

6. Tools Return Results

A tool usually returns some form of result. browser_navigate might return "Page loaded successfully"; query_database might return product_id = 101, name = Laptop, price = 999. The AI can then use the result as a new observation.

7. Tools Are Important for Agentic AI

Tools are what allow AI agents to move beyond simply generating text.

8. What Is an MCP Resource?

An MCP Resource represents information or context that can be accessed through an MCP server. Think: Resource = Data or Context. Examples could include files, documents, configuration, schemas, logs, repository content, test specifications and application metadata. A resource is generally something the AI can retrieve and use as context.

9. Resource Example for an SDET

Suppose your test environment has a Product Service Specification. The AI may need to understand required fields, business rules, validation rules, API contracts and expected behavior. That specification could conceptually be available as a resource.

The AI can retrieve the information and use it for reasoning.

10. Tools vs Resources

This distinction is critical. If the AI needs to read requirements, that might be represented as a Resource (Acceptance Criteria). If the AI needs to test the browser, that might use a Tool (browser_click).

11. Easy Example

Imagine an AI testing agent. It first needs the test specification (Resource: Read Acceptance Criteria), then it needs to interact with the browser (Tool: Open Application), then another tool (Click Add Product), then maybe another resource (Read API Schema). The agent combines information and actions.

12. Resource vs Tool Example with Database

Suppose the system exposes database schema information. A resource might be the products table schema (product_id, name, price, quantity, in_stock) — that provides context. But executing a query could be exposed as a tool: query_database.

This is a good way to understand the difference.

13. Resource Identifiers

Resources typically need a way to be identified. The exact format depends on implementation, but the important idea is that a client can identify and retrieve available information.

14. Why Resources Matter for AI

An AI model cannot automatically know your company's requirements, test plans, database schema, internal architecture, business rules or application documentation. Resources can provide that context. Instead of relying only on general model knowledge, the system can supply enterprise context.

That makes the AI more useful for real enterprise testing.

15. What Is an MCP Prompt?

An MCP Prompt is a reusable prompt template exposed by an MCP server. Think: Prompt = Reusable Instruction. A prompt can help standardize how certain tasks should be performed.

Instead of rewriting the instruction every time, it could be available as a reusable prompt.

16. Prompt Example for SDET

Imagine an MCP server exposes a prompt called analyze_test_failure. Conceptually, it could contain instructions like:

Now the AI application can surface or use that standardized workflow.

17. Why Prompts Are Useful

Prompts help teams create consistency. Without standardized prompts, Tester A says "Check why this failed", Tester B says "Analyze this issue", and Tester C says "Debug it" — each instruction could produce different approaches. With a reusable Failure Analysis Prompt, everyone can follow a more standardized process. This can be especially useful in enterprise Quality Engineering.

18. Tools vs Resources vs Prompts

• Tool — perform an action.
• Resource — access data/context.
• Prompt — provide reusable instructions.

19. Playwright Example Using All Three

Suppose the goal is: "Verify the Product Creation feature." The workflow could use all three MCP capabilities.

This is a complete MCP-enabled workflow.

20. Complete Example

21. MCP Server Capability Discovery

The MCP client can discover what capabilities a server exposes. This means an AI environment can learn what is available instead of assuming every MCP server provides the same capabilities.

22. Tool Discovery

A response might conceptually include browser_navigate, browser_click, browser_type and browser_snapshot. Then the AI knows which actions are available.

23. Resource Discovery

Possible results: Application Specification, Test Data Guide, Database Schema, API Contract. The AI can then request one of those resources.

24. Prompt Discovery

The client may also discover reusable prompts.

• Analyze Test Failure
• Generate Test Scenarios
• Review Acceptance Criteria
• Create Defect Summary

25. Enterprise Testing Example

Imagine your enterprise MCP ecosystem contains several servers. Different servers may expose different capabilities.

26. Playwright MCP Server

Could expose tools such as Navigate, Click, Type, Inspect Page and Take Snapshot. Main purpose: Browser Interaction.

27. GitHub MCP Server

Could expose capabilities related to repository content, issues, pull requests, code search and file access. Depending on the implementation, some may be tools and some may be resources. For example: Read repository file → Resource or retrieval capability; Create issue → Tool.

28. Database MCP Server

29. Jira MCP Server

Now an agent could potentially do:

30. Prompts in Enterprise Testing

A team might create standardized prompts such as: Review Requirement, Generate Test Scenarios, Analyze Automation Failure, Perform Root Cause Analysis, Generate Defect Report, Generate Test Summary. This can provide consistent engineering workflows across the organization.

31. Example: Requirement Review Prompt

An SDET could use this repeatedly for new stories.

32. Example: Test Design Prompt

This helps standardize AI-assisted test design.

33. Example: Failure Analysis Prompt

This can be very useful for SDET workflows.

34. The AI Can Combine Tools

An agent is not limited to one tool. Together, a sequence of actions creates a workflow. The AI determines the sequence based on the goal and observations.

35. The AI Can Combine Multiple Servers

The AI may orchestrate multiple capabilities as part of one testing objective.

36. Example End-to-End Agentic Workflow

Goal: "Verify that an administrator can create a new product."

1. RESOURCE — Read PRODUCT-101 acceptance criteria
2. PROMPT — Generate test strategy
3. TOOL — Open QA application
4. TOOL — Login as administrator
5. TOOL — Create product
6. TOOL — Query database
7. AI — Compare expected vs actual
8. If failure occurs: PROMPT — Analyze Failure
9. If confirmed defect: TOOL — Create Jira Bug

37. Tools Need Proper Descriptions

A tool should be clearly described. Imagine two tools: "run" and "query_qa_database_read_only". The second is far easier for the AI to understand safely. A good tool definition should communicate what the tool does, what input it needs, what it returns and what restrictions apply. Good tool design improves reliability.

38. Tool Inputs Should Be Clear

This helps the AI understand exactly how the tool should be used.

39. Avoid Overly Powerful Tools

Tool design is also a security concern.

The more specific the capability, the easier it can be to control.

40. Least Privilege Applies to Tools

Suppose a testing agent only needs to validate test data. It should not automatically receive DROP DATABASE, DELETE production records or Modify production users. Instead, Read QA data may be sufficient. Use: Minimum Necessary Capability.

41. Resource Security

Resources may contain sensitive data too — production secrets, customer PII, passwords, internal architecture or API keys should not automatically be exposed to every AI workflow. Resources should follow proper authentication, authorization, data classification, access control and audit logging.

42. Prompts Also Need Governance

Prompts may sound harmless, but they can influence agent behavior.

Prompt design is part of responsible agent architecture.

43. Tools Are Not the Same as APIs

A tool may internally call an API.

44. Resources Are Not Necessarily Files

A resource might be a file, but it does not have to be. A resource could conceptually represent a document, configuration, schema, dynamic application information, repository content or reference information. The important characteristic is: it provides information/context.

45. Prompts Are Not Tools

A prompt provides instructions. A tool performs an action.

Very different responsibilities.

46. Resource + Prompt + Tool Together

This is the most important combination.

47. Traditional Automation vs MCP Capabilities

This does not mean deterministic automation disappears. They can work together.

48. Deterministic Playwright Still Matters

Traditional Playwright remains ideal for regression testing, CI/CD, repeatable validations, release gates and stable business flows. MCP and agentic workflows can complement this by supporting exploratory testing, requirement analysis, failure investigation, dynamic test planning and cross-system orchestration.

49. Enterprise Hybrid Model

50. Realistic SDET Example

Suppose your Product Service application supports UI, REST, GraphQL, WebSocket, PostgreSQL and Kafka. An agent may need:

This is a powerful way to structure AI-assisted Quality Engineering.

51. Architecture Example

52. Agent Reasoning Loop

Resources can supply context during this cycle. Prompts can help structure reasoning. Tools execute actions.

53. Common Interview Question

What is the difference between MCP Tools, Resources, and Prompts? A strong answer: MCP Tools expose executable capabilities that allow an AI application to perform actions. MCP Resources expose information or context that can be retrieved and used by the AI. MCP Prompts provide reusable prompt templates that help standardize instructions or workflows.

54. Another Interview Question

Can an MCP server expose all three? Yes. An MCP server may expose Tools, Resources and Prompts depending on what the server is designed to provide. Some servers may expose mostly tools; others may primarily provide information.

55. Another Interview Question

Does the AI model execute the tool directly? Not exactly.

The model decides when and why to use the capability. The external tool performs the underlying action.

56. Another Interview Question

Can MCP Tools call REST APIs? Yes.

MCP does not replace the application's API. It may provide an AI-friendly capability layer over it.

57. Easy Analogy

Imagine an engineer in a workshop.

58. MCP Capability Cheat Sheet

59. What You Should Memorize

60. Final Takeaway

Together, these capabilities allow AI applications to move from simple conversation toward useful engineering workflows.

That is one of the foundations of MCP-based Agentic Quality Engineering.