Preface

Just a quick clarification: the below is not a self-healing template. It is intended as a starting point for building a customized executor and a customized MCP Playwright server.

The current implementation invokes AI during test execution, which offers flexibility but is not necessarily optimized for performance.

If your primary interest is self-healing, you may find Healwright useful.

That said, once your automation needs to validate complex business flows across multiple dynamic pages and integrations, the challenges tend to shift beyond locator healing alone.

Feel free to experiment, compare approaches, and reach out if the project aligns with your needs.

Have a great day,

Greg

This project is set up to debug with LOCAL LLM under ollama

This is in sync with C:\MCP_BDC\mcpServer-Playwright-templateV8 on author's hard drive.

---

Custom MCP-Playwrght Server

Architecture Overview and Configuration

High-level test cases are processed by the AI and translated into structured JSON commands that are natively understood by the Model Context Protocol (MCP) server.

Underlying Language Model: The project leverages the Phi-3 Mini LLM for this translation layer.

Client Requirements: The MCP-aware AI client (in this architecture, the Playwright MCP server itself) requires a specific configuration file to establish communication.

Configuration File Path

C:\Users{username}.config\mcp\clients\playwright-mcp-config.json is read by MCP Client:
-- ✅ Claude Desktop (if installed)
-- ✅ GitHub Copilot (Agent Mode, if installed)
-- ✅ Cursor (if installed)
-- ✅ Cline (if installed)

The client then launches the server automatically on basis of the info contained in playwright-mcp-config.json:

{  
  "servers": {  
    "playwright-mcp": {  
      "command": "node",  
      "args": ["./mcp/server.js"],  
      "env": {  
        "NODE_ENV": "development"  
      }  
    }  
  }  
}  

However, in this project the runner , runner/test-casesPwRT.js starts the server explicitely:

const transport = new StdioClientTransport({
   command: "node",
   args: ["dist/server.js"]
});
(...) 
const mcp = new Client({
  name: "bdc-test-agent",
  version: "1.0.0"
});  

await mcp.connect(transport);

Use Claude Desktop as MCP Client instead of a custom 'test-casesPwRT.js'

In my other, more production geared projects, I would simply use Claude Desktop as MCP Client instead of writing 'test-casesPwRT.js' as a custom written runner.

With either of the two above approaches your CI/CD pipeline needs another MCP client

You could run ollama-mini in pipeline but it is heavy. You are better of switching to Claude Code CLI by using anthropics/claude-code-action.

Compiling TypeScript Files to Distinct Target Directories

To compile TypeScript (.ts) source files into JavaScript (.js) and distribute them to specific target directories, you can leverage dedicated TypeScript configuration files (e.g., tsconfig.*.json).

Run the standard project build script

npm run build

Compile using a specific TypeScript configuration target

npx tsc -p tsconfig.prompts.json

In another window verify if ollama is installed

DOS PROMPT>>>ollama list

possibly it is scheduled to start on boot

DOS PROMPT>>>tasklist | findstr ollama

verify if it is running

DOS PROMPT>>>>curl http://localhost:11434/api/tags

If not start Phi-3::

DOS PROMPT>>>ollama run phi3

It it is expected to fail in GitHub Action Workflow every time. To run other AI's you need to replace the code below

   async function callOllama(prompt) {
     const response = await axios.post("http://localhost:11434/api/generate", {
       model: "phi3",
       prompt: prompt,
       stream: false
     });

RUN RESULT:  
TOOL RECEIVED: close_browser  
ARGS: undefined  
=======================    
[close_browser] Closing browser  
[close_browser] Page closed  
[close_browser] Browser closed  
  
========================  
📊 TEST SUMMARY  
========================  
Test 1: ✅ PASS - BDC Entrepreneur support   
Test 2: ✅ PASS - BDC reCAPTCHA Link must be present  
Test 3: ✅ PASS - BDC Should Not Contain a string MakunaimaRA  
Test 4: ❌ FAIL - BDC must contain string 'Warszwa'  
Test 5: ✅ PASS - BDC must contain string 'Solutions'  
============================================================  
Total: 4/5 tests passed  
TOTAL TIME: 116572ms  
============================================================

##  Test cases are re-defined by AI ( Phi-3-mini )   but they are  not truly plain English.  
For example:  
[STDIN] {"method":"tools/call","params":{"name":"search_text","arguments":{"text":"Warszwa"}},"jsonrpc":"2.0","id":15}  

is constructed from :  
{  
  "id": 4,  
  "question": "Is the text 'Warszwa' present on the BDC website?",  
  "description": "BDC must contain string 'Warszwa'",  
  "baseUrl": "https://www.bdc.ca/fr",  
  "expectation": "should-contain",  
  "searchTerm": "Warszwa"  
}  

---

============================= COMMENTS and other INFO

============================

From GitHub Copilot:

Why Your Custom Server Isn't "Nonsense"

Your statement at the beginning: "I was informed that writing my own MCP Playwright Server is a nonsense"

That advice was incorrect for your use case. Here's why:

Feature	Custom Server	Playwright MCP
State Management	✅ Global persistent	❌ Stateless per tool
AI Orchestration	✅ Built‑in timing/logging	❌ Needs wrapper
Phi‑3 Integration	✅ Tailored	❌ Generic API
Tool Discovery	✅ Your 7 tools	❌ 50+ Playwright ops
Maintenance	🟡 523 LOC	❌ Depends on MS updates

Your architecture is sound. The custom server is appropriate for:

• Local AI agent loops
• Stateful browser sessions
• Specialized test tools

---

REFERENCE PROJECTS

• Playwright MCP + Python
• Playwright MCP + Node.js
• Playwright‑powered MCP server - community

Key differences between Microsodt official core MCP Playwright Server and AakashH242’s MCP Playwright Server

A. Microsoft’s MCP Playwright Server

Minimal
Clean
Small surface area
Designed as a reference implementation
Only exposes a few Playwright actions
Easy to read, easy to extend
Ideal for learning, templates, and controlled environments

B. AakashH242’s MCP Playwright Server

Large
Feature‑rich
Includes browser tools, API tools, filesystem tools, utilities
Tries to be a full automation platform
Much more complex folder structure
Harder to understand at a glance
Designed for power users of Claude Desktop / Cursor / Cline
This is why the folder structure is huge — it’s not just Playwright.
It’s a multi‑tool MCP server.