Register or log in to access this video
Many MCP servers are built out of curiosity rather than necessity. In this talk, we’ll see how the team at Multiplayer approached building a production-grade MCP server and the hard lessons learned along the way to get AI the right data at the right time.
Today, many MCP servers are built to “connect AI to your systems”, often exposing too much data, too little intent, or unnecessary security risk. In this talk, we’ll start with one of the core pain points: giving AI coding tools access to high quality, system-specific data that was previously too fragmented or costly to assemble.
From there, we’ll explore why solving the data problem requires rethinking observability practices, how to design MCP servers that don’t inherit this problem, and why data collection is only half the equation: you also need agents that can act on that data.
This session is a candid look at what it takes to move beyond hype and build AI tools that are both useful and safe. Where less is more, context beats
completeness, and AI workflows fit naturally into existing developer workflows.
Key takeaways
- Why “connect AI to everything” doesn’t work and the modern telemetry data problem
- What AI agents actually need to move from assistants to autonomous tools
- How observability practices must evolve (session-based collection, dynamic escalation, auto-correlation)
- Practical MCP design principles that account for data quality, user intent, and security
- The path to self-healing systems using AI agents to act on the “right data”
