← back
Bending a Public MCP Server Without Breaking It — Nimrod Hauser, Baz
Original: Bending a Public MCP Server Without Breaking It — Nimrod Hauser, Baz
Takeaway
You can dramatically improve third-party MCP tool behavior by rewriting descriptions, scoping, and wrapping them — without forking the upstream server.
Summary
- Baz's Nimrod Hauser uses Playwright's MCP server as a running example of third-party tools degrading agent performance.
- Diagnoses common failures: bad tool descriptions, over-broad scopes, missing schemas, and tools that overlap or conflict.
- Presents a five-best-practice framework to 'bend' public MCP servers — refine descriptions, prune tools, wrap with adapters, add typing, and observe usage.
- Demonstrates the fixes resolving an agent that initially 'caught fire' from a busted public MCP server.
mcpagentstool-use
Original description
Public MCP servers often look ready-to-use, until the reality of production hits. You might find your agents ignoring perfectly good tools, unwanted side-effects exhausting your container's disk space, or worse, security concerns like multi-tenant leaks wreaking havoc. What begins as a ""simple integration"" can quickly become a source of friction and unexpected failure. In this talk, we'll share a hands-on guide to adapting third-party MCP servers for real-world applications. You'll learn practical processes to identify friction points and strategies to modify MCP servers so they integrate seamlessly with your specific agents and architecture. Real-world lessons, trade-offs, and production-tested solutions included. Using a concrete example, we'll walk through the journey of transforming a brittle setup into production-ready infrastructure. We'll cover editing tool definitions, optimizing agentic context, and layering deterministic validations—all while preparing for scale. This iterative debugging process will provide you with a repeatable framework to make any MCP integration resilient, secure, and production-ready. Nimrod Hauser - Founding Software Engineer, Baz Nimrod is a Principal Engineer at Baz, building AI-powered code review agents. A “jack of all trades” across backend, data engineering, and data science, he has worked at the intersection of software and data throughout his career. He began as a data analyst in the military, helped lay the foundations of Salesforce’s Einstein platform, and later became the first data scientist at cybersecurity startup BlueVoyant. He went on to lead data and architecture at Solidus Labs in the crypto-regulation space before joining Baz. Nimrod thrives on building systems from scratch and turning ideas into scalable products. Socials: https://www.linkedin.com/in/nimrod-hauser-03776a31/ https://x.com/NimrodHauser Slides: https://prezi.com/view/TSBwBXLNcXzzWrLbRiit/?referral_token=4jzLrblnB3FN