Bending a Public MCP Server Without Breaking It
summary
Shows a concrete engineering path: starts with Playwright MCP server offering 21 generic tools that cause a vanilla run to fail, then iterates through five practical moves—curate tools, wrap and rewrite descriptions, add deterministic guardrails (path validation for screenshots), compose an evidence tool, and pull login out of agentic flow using JWTs. Demonstrates a spec‑reviewer built end‑to‑end, measures tool count drop and a failing→passing run, and shares code patterns and tradeoffs useful for immediate reuse.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 4:32Agentic tools challenges
- 8:08Vanilla tools baseline
- 14:26Optimizing third-party tools
- 15:50Curating tools context
- 18:22Tailored tool descriptions
- 22:55Guardrails for security
- 28:58Composing tools
- 32:57Tools as functions
- 36:52Demo and results
- 39:21Summary
related talks
43mServers Beyond 101 Good Practices, Design Choices and Their Consequences
17mBuilding Production-Ready MCP Servers: From AI Agents to Secure Enterprise APIs.
12mPrompting LLMs to use Tools and Resources with MCP
24mThese Aren't the Tools You're Looking For: MCP Security Awakens
29mMCPs are Boring (or: Why we are losing the Sparkle of LLMs)
02h 49mMCP Servers: Best Practices, Design Choices and Their Consequences
01h 31mBuild AI-Powered Automation Tools: MCP Workshop for Developers & Engineers
31mWriting Model Content Protocol (MCP) Servers in Clojure
