How to connect AI to your testing workflow with MCP

MCP (Model Context Protocol) lets your AI assistant connect directly to your test management tool. Instead of copy-pasting feature descriptions into ChatGPT and manually creating test cases from the output, the AI reads your existing test repository, understands your structure, and creates scripts in place. Setup takes about 10 minutes.

If that sounds too good, here's the catch: you still need to review what the AI generates. AI is good at producing a first draft of test cases from requirements. It's not good at understanding your business domain, your users' weird habits, or the specific bug that only happens on Safari with a VPN. The human reviews, adjusts, and executes. The AI handles the mechanical parts.

For a deeper look at what MCP is and how the protocol works, see our MCP explainer. This article is the practical setup guide.

Why the copy-paste workflow is broken#

Before MCP, using AI for testing looked like this:

1. Copy a feature description or user story
2. Paste it into ChatGPT/Claude with a prompt like "generate test cases for this"
3. Read the output
4. Manually create each test case in your test management tool
5. Realize the AI didn't know about your existing tests, so some cases are duplicates
6. Edit, merge, reorganize

It works, technically. But you lose context at every step. The AI doesn't know what you've already tested. It doesn't know your tagging convention. It doesn't know that your team structures tests with headers for feature areas and child items underneath. Every interaction starts from zero.

Gartner predicted that by 2026, 30% of enterprises would use AI-augmented testing. The bottleneck isn't the AI's ability to generate test cases. It's the integration between AI and the tools where test cases actually live.

What MCP changes#

MCP is an open protocol that Anthropic released in November 2024. By March 2025, OpenAI had adopted it. By December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, with support from OpenAI, Google DeepMind, and Microsoft. It's now the standard way AI connects to external tools.

The architecture is simple. Your AI client (Claude Desktop, VS Code, Cursor) connects to an MCP server. The MCP server connects to your test management tool's API. The AI can then read your projects, list scripts, create test items, start runs, and submit results, all through natural language commands.

The difference from the copy-paste approach: the AI sees your entire test repository. It knows what's already covered. When you ask it to generate test cases for a new feature, it can check existing scripts and avoid duplicates. When you ask it to find coverage gaps, it can scan your whole test suite and point to areas with thin coverage.

Step-by-step setup#

Step 1: Get your MCP server details#

Your test management tool needs to provide an MCP server. In TestRush, MCP is built in from day one. You'll need two things from your tool:

• The MCP server URL (usually something like https://mcp.yourtool.com/sse)
• An API key for authentication

In TestRush, you find these in your project settings. The API key is tied to your account, so the AI operates with your permissions.

Step 2: Configure your AI client#

Each AI client stores MCP server configs in a JSON file. The location depends on the client:

Claude Desktop: The config file is at ~/Library/Application Support/Claude/claude_desktop_config.json on Mac or %APPDATA%\Claude\claude_desktop_config.json on Windows.

VS Code / Cursor / Windsurf: These editors use a .mcp.json file in your project root or workspace settings.

The config structure is similar across clients:

{
  "mcpServers": {
    "testrush": {
      "url": "https://mcp.testrush.app/sse",
      "headers": {
        "x-api-key": "your-api-key-here"
      }
    }
  }
}

Save the file, restart your AI client, and the connection should be active.

Step 3: Verify the connection#

Ask your AI a simple question to confirm it can reach your test management data:

"List all my projects in TestRush"

If the connection works, the AI will return your project names and IDs. If it fails, check the URL and API key. The most common issue is a typo in the config file or a missing restart of the AI client.

Step 4: Do something real#

Now try a real task. Give the AI a feature description and ask it to create test cases:

"Here's the spec for our new password reset flow: [paste spec]. Create a test script called 'Password Reset' in my project, with sections for happy path, error handling, and edge cases."

The AI will create a script with headers and child items, using the structure your tool supports. In TestRush, that means headers for each section and testable items underneath, tagged appropriately.

Review what it creates. The happy path cases will probably be solid. The edge cases might need your domain knowledge. "What happens when the user's email is associated with a deactivated account?" is the kind of thing AI misses unless your spec mentions it.

What you can do with MCP connected#

Once the connection is live, here are the workflows that actually save time:

Generate test scripts from requirements. Paste a PRD, user story, or feature spec into your AI conversation. Ask it to create a test script. The AI produces a structured set of test cases, already in your tool, ready to run. See our guide on AI test case generation for prompt tips.

Find coverage gaps. Ask: "Look at my existing test scripts and tell me what areas have thin coverage." The AI scans your test repository and identifies feature areas with few test cases or scenarios that aren't covered (boundary conditions, error states, permissions).

Build regression suites from existing scripts. Ask: "Create a regression suite for the checkout feature by pulling the most important test cases from my existing scripts." The AI reads your scripts, picks high-priority items, and assembles them into a new script tagged "regression."

Analyze test run results. After a run, ask: "What patterns do you see in the failed test items from run #47?" The AI can identify clusters of failures (all failures in the payment module, all failures related to timeouts) and suggest where to focus debugging.

Which AI clients support MCP#

The protocol is client-agnostic, but these are the ones with production-ready MCP support as of early 2026:

• Claude Desktop and Claude Code -- native MCP support, the smoothest experience since Anthropic created the protocol
• VS Code with Copilot or Claude extensions -- via .mcp.json in your workspace
• Cursor and Windsurf -- built-in MCP configuration
• ChatGPT desktop app -- OpenAI adopted MCP in March 2025
• Local LLMs via Ollama -- if you're running models locally, MCP-compatible clients can connect them to your tools

The setup process is nearly identical across clients. The config file location changes, but the JSON structure is the same.

Security considerations#

A few things to keep in mind:

Data stays in your account. MCP is a connection protocol. Your test data lives in your test management tool's database. The AI reads and writes through the API, but your data isn't copied to a third-party server.

The AI doesn't store your tests. During a conversation, the AI processes your test data in memory. When the conversation ends, that context is gone. The next conversation starts fresh (unless your AI client has memory features, which are separate from MCP).

Permissions match your account. The AI operates with whatever permissions your API key grants. If your key has read-only access, the AI can read but not create. Set the minimum permissions you need.

Local config files hold your keys. The MCP config file on your machine contains your API key. On shared computers, this is a risk. On your personal machine with full-disk encryption, it's fine.

Common mistakes#

1. Trusting AI output without review. AI generates good first drafts. It misses domain-specific edge cases, business rules that aren't in the spec, and the bug that only your senior tester knows about. Always review before running.

2. Overcomplicating the setup. MCP config is one JSON file with a URL and a key. If you're spending more than 15 minutes on setup, something is wrong. Check the basics: correct URL, valid API key, client restarted.

3. Using AI for execution instead of generation. MCP connects AI to your test management tool. The AI generates scripts, finds gaps, analyzes results. It doesn't click through your app and check if buttons work. That's still a human job (or a Selenium job).

4. Not having test cases to begin with. AI works best when it can see your existing test structure and build on it. If your test repository is empty, the AI has no context. Write some test cases first, even a small set, before connecting AI.

FAQ#

What AI models work with MCP?#

Claude has native support through Claude Desktop and Claude Code. GPT works through the ChatGPT desktop app and API adapters. Gemini and local models via Ollama also support MCP through compatible clients. The protocol is open and model-agnostic.

Is my data safe?#

Your test data stays in your test management tool. MCP is a connection protocol, not a data storage layer. The AI reads and writes through your tool's API during your session. API keys are stored locally on your machine. No test data is permanently stored by the AI.

Does MCP cost extra?#

MCP is an open protocol with no licensing fees. You pay for your AI subscription and your test management tool subscription. TestRush pricing includes MCP access at every tier, no add-on charges.

Can I use MCP with a self-hosted LLM?#

Yes. If you run a local model through Ollama or a similar setup, you can connect it to MCP servers. The AI runs on your hardware, and MCP connects it to your tools. Your data never leaves your network.

Ready to connect AI to your testing workflow? Start your free trial or see MCP in action in the demo.