QA MCP Server — Full Guide
Everything you need: what MCP is, how to connect this server to your IDE or language, and how to use all 15 tools effectively. No prior MCP experience required.
Contents
1. What is MCP?
MCP stands for Model Context Protocol — an open standard developed by Anthropic that lets AI assistants (like Claude) talk to external tools, APIs, and services in a structured way.
Think of it like this: before MCP, if you wanted Claude to check your test results, you had to copy-paste them into the chat. With MCP, Claude can call a tool directly — without leaving your editor.
The analogy
MCP is to AI what REST APIs are to web apps. It gives AI a standardised way to call external capabilities — generate test cases, parse a JUnit report, evaluate a prompt — and get structured results back. The AI can then reason over those results, ask follow-up questions, or chain multiple tool calls together.
Which tools support MCP?
2. What is this QA MCP Server?
This is a purpose-built MCP server that exposes 15 AI-powered QA & SDLC tools to any AI client or language. Instead of opening a web app, you can call these tools directly from your IDE or scripts while you work.
BDD + plain from feature descriptions
Postman, pytest, REST Assured, Karate
Manual → Playwright, Selenium, Cypress
JUnit / Allure → root cause insights
Raw description → JIRA-ready report
Coverage score + missing scenarios
Feature → person-day breakdown
Full plan with risks + sign-off
Role + seniority → targeted Qs
Incident → blameless postmortem
All tools are language-agnostic — the MCP protocol means it does not matter if you code in Python, Java, TypeScript, or Go. The server handles everything.
3. Connecting to Your IDE
Pick your editor and follow the one-time setup. After this, the tools are available in every conversation or code session.
Claude Desktop
- 1
Open config file
Mac: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%/Claude/claude_desktop_config.json
- 2
Add this block
{ "mcpServers": { "qa-intelligence": { "url": "https://qa-mcp-server-theta.vercel.app/api/mcp" } } } - 3
Restart Claude Desktop
Quit and reopen. You should see the QA tools appear when you click the ⚡ icon in a conversation.
Cursor
Create or edit .cursor/mcp.json in your project root:
{
"servers": {
"qa-intelligence": {
"type": "http",
"url": "https://qa-mcp-server-theta.vercel.app/api/mcp"
}
}
}Reload the window. In any Cursor chat, ask Claude: "Generate BDD test cases for the login feature" and it will call the tool automatically.
VS Code Copilot
Create .vscode/mcp.json:
{
"servers": {
"qa-intelligence": {
"type": "http",
"url": "https://qa-mcp-server-theta.vercel.app/api/mcp"
}
}
}4. Using from Python, TypeScript, Java
Any language with an HTTP client can call the tools. Install the MCP SDK for the best experience, or use the Playground REST API for a simpler direct call.
Python — via MCP SDK
pip install mcp
import asyncio
from mcp.client.streamable_http import streamablehttp_client
from mcp import ClientSession
async def main():
async with streamablehttp_client(
"https://qa-mcp-server-theta.vercel.app/api/mcp"
) as (read, write, _):
async with ClientSession(read, write) as session:
await session.initialize()
# Generate BDD test cases
result = await session.call_tool(
"generate_test_cases",
{
"feature_description": "User can reset password via email OTP",
"format": "bdd",
"negative_cases": True
}
)
import json
print(json.loads(result.content[0].text))
asyncio.run(main())TypeScript / Node.js — via MCP SDK
npm install @modelcontextprotocol/sdk
import { Client } from '@modelcontextprotocol/sdk/client/index.js'
import { StreamableHTTPClientTransport } from '@modelcontextprotocol/sdk/client/streamableHttp.js'
const client = new Client({ name: 'my-qa-pipeline', version: '1.0' })
const transport = new StreamableHTTPClientTransport(
new URL('https://qa-mcp-server-theta.vercel.app/api/mcp')
)
await client.connect(transport)
// Parse failed test results
const result = await client.callTool('parse_test_results', {
results: junitXmlContent,
format: 'junit-xml'
})
const analysis = JSON.parse(result.content[0].text)
console.log('Failed tests:', analysis.failures)Java — via Playground REST API
The Java MCP SDK is in early development. Use the Playground API endpoint directly — it is a plain HTTP POST:
// Using OkHttp (or any HTTP client)
OkHttpClient client = new OkHttpClient();
String requestBody = """
{
"passcode": "YOUR_PASSCODE",
"tool": "generate_bug_report",
"args": {
"description": "Checkout fails silently when cart has 5+ items",
"environment": "Chrome 120, macOS, production",
"severity": "high"
}
}
""";
Request request = new Request.Builder()
.url("https://qa-mcp-server-theta.vercel.app/api/playground")
.post(RequestBody.create(requestBody,
MediaType.get("application/json")))
.build();
try (Response response = client.newCall(request).execute()) {
System.out.println(response.body().string());
}5. Tool Walkthrough: All 15 Tools
Here is when to reach for each tool and what it gives you back.
generate_test_casesgenerate_api_testsconvert_to_automationgenerate_test_dataparse_test_resultsgenerate_bug_reportfind_test_gapsestimate_qa_effortgenerate_test_plangenerate_interview_questionsevaluate_promptevaluate_rag_outputanalyze_qa_candidategenerate_postmortemanalyze_logs6. The Playground (no IDE needed)
The Playground is a web interface where anyone can try all 15 tools without installing anything or connecting an IDE. Perfect for demos, onboarding, or exploring capabilities.
- 1
Enter the passcode
Request the passcode from the server owner. It is shared separately to prevent public abuse.
- 2
Pick a tool
Choose from the left sidebar. Every tool has a "Load Example" button — hit that first to see it in action with a pre-filled, realistic input.
- 3
Fill in your inputs
Replace the example with your own content. Required fields are marked with ★. Optional fields add precision but are not needed.
- 4
Click Run
The tool calls the AI and returns structured JSON. Use "Copy JSON" to grab the output for your pipeline or documentation.
7. Common Patterns & Workflows
→ Feature testing workflow (start to finish)
- 1.
generate_test_cases— get BDD scenarios from the user story - 2.
find_test_gaps— cross-check against any existing tests - 3.
generate_test_data— get realistic data for the test cases - 4.
convert_to_automation— convert the top priority cases to Playwright
→ CI/CD failure triage
- 1.
parse_test_results— understand what failed and why - 2.
analyze_logs— correlate with application logs if tests are environment-sensitive - 3.
generate_bug_report— raise the top failures as structured JIRA tickets
→ Sprint planning
- 1.
estimate_qa_effort— get an effort breakdown for the sprint scope - 2.
generate_test_plan— produce a test plan for the sprint - 3.
find_test_gaps— identify what regression to prioritise
→ New API → tests in 3 calls
- 1.
generate_api_tests— get Postman collection + test code from the spec - 2.
generate_test_data— get request payloads for edge case testing - 3.
evaluate_prompt (optional)— if the API uses an LLM, evaluate its prompt quality
8. Tips & Troubleshooting
Tip: All outputs are JSON
Every tool returns structured JSON, not prose. This means you can use the output programmatically — pipe it into a script, save it to a file, or feed it into another tool.
Tip: Chain tools in sequence
The output of one tool can become the input of the next. For example: use generate_test_cases, then pass the test_cases array to find_test_gaps as existing_tests.
Tip: More input = better output
These tools are AI — the more context you provide, the more accurate and specific the output. A 3-sentence feature description produces generic cases. A 2-paragraph description with constraints produces precise, relevant cases.
Avoid: Do not paste sensitive data
The tools call an AI API — do not paste production credentials, PII, or confidential customer data into any field. Use anonymised or synthetic data for testing.
Avoid: Treat output as a draft, not final
AI-generated test cases, effort estimates, and bug reports are starting points. Review and adjust before using in production workflows. The tools dramatically reduce time-to-first-draft, not thinking time.
Avoid: "Connection refused" or 404
The MCP endpoint URL must match exactly. Make sure you are using /api/mcp (not /api or /mcp). If self-hosting locally, the server must be running on the expected port.
Ready to try it?
Use the Playground for zero-setup access, or connect your IDE for the full experience.