How We Built Our First MCP Server: Lessons From the Trenches at Hud
summary
Hud builds a production MCP server that connects live runtime signals to code by computing a function identity (no source stored) and matching it to the editor’s unit of work; uses Anthropic’s SDK and MCP inspector debugger plus Tree‑Sitter for fast AST matching; discovers the server only becomes useful after renaming tools to actionable labels like get endpoint error breakdown; handles scale (an endpoint example gets ~350 million calls/day) and embeds rules to gate deploys and error fixes.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 2:52Model Context Protocol
- 5:47First MCP server setup
- 7:53Production vs local code
- 13:11User-focused tools
- 15:39Rules for AI behavior
- 16:22Runtime context in workflows
- 19:29Production issues with MCP
- 22:00Key lessons learned
- 24:11Context engineering
- 25:56Final thoughts on MCP
related talks
28mBuilding MCP servers with VS Code — Level up your MCP
17mMCP Is Not Good Yet
55mYour MCP Server is Bad (and you should feel bad)
25mMaking AI Personal: Building Context-Rich AI Tools with MCP
15mFull Spec MCP: Hidden Capabilities of the MCP spec
17mThe State of MCP observability: Observable.tools
01h 29mWorkshop: Build Your Testing Sidekick- Custom Tools with Model Context Protocol
25mCruising the AI Highway: Standardized MCP Deployment and Governance With Backstage