web
Stunt Double MCP Server
Connecting AI assistants to automated UX validation through the Model Context Protocol.
Year
2026
Technologies
JavaScript, AI, AI Agents, MCP, Claude, Cursor, UX Design
I built the Stunt Double MCP Server to bridge the gap between AI assistants like Claude and automated UX testing. By exposing Stunt Double's persona-driven AI agents through the Model Context Protocol (MCP), developers can now validate user journeys directly from their IDEs or AI chat interfaces.
The Challenge
Modern web development moves incredibly fast, making it difficult to catch UX friction points before they reach real users. While traditional end-to-end testing catches functional bugs, it fails to evaluate the subjective "feel" and usability of a product. We needed a way to deploy realistic AI user personas that could navigate interfaces and report friction.
More importantly, we needed developers to trigger and interact with these personas effortlessly. Context switching between code editors, testing dashboards, and AI assistants creates unnecessary friction in the developer experience.
Approach
I implemented an MCP service that acts as the translation layer between LLM-powered development tools and the core Stunt Double engine. Instead of building a heavy Node.js wrapper in this specific repository, I designed it as a lightweight configuration and rules engine containing plugin manifests, skills, agents, and Cursor rules.
The actual MCP endpoint is hosted securely on the main Stunt Double infrastructure (https://app.stuntdouble.io/api/mcp), keeping the client-side integration virtually zero-dependency. I optimized the integration paths for Claude Code, Claude Desktop, and Cursor/Windsurf. This architectural decision ensures that adding the server requires nothing more than a simple configuration block or a single command-line execution.
Key Features
- Seamless IDE Integration: Native support for Cursor and Windsurf via
mcp.jsonconfiguration blocks. - Claude Ecosystem Support: Instantly add the server to Claude Desktop and Claude Code CLI via simple HTTP transport.
- Agent Configuration Engine: Houses specialized skills, actors, and rules to simulate realistic user behaviors on web apps.
- Lightweight Architecture: Zero-dependency client-side repository focusing purely on MCP configuration and LLM schemas.
- Automated Journey Validation: Programmatically deploy AI personas to test workflows and surface friction points before deployment.
Results
The MCP server enables development teams to integrate Stunt Double directly into their existing AI workflows, dramatically reducing the friction of setting up UX testing. Developers can now simply ask their AI assistant to "run a usability test on the new checkout flow," triggering our personas to execute automated workflows against the app without ever leaving the code editor.
What I Learned
- Model Context Protocol (MCP): Gained deep insights into crafting effective MCP configurations, tools, and schemas that LLMs can naturally understand and utilize without hallucinating.
- AI Tooling Ecosystems: Mastered the nuances of integrating with different AI environments, balancing the specific transport and parsing requirements of Cursor, Claude Desktop, and CLI tools.
- Decoupled Architecture: Discovered the immense UX benefits of separating the configuration repository from the hosted execution engine to simplify user onboarding and eliminate package management overhead.
