Engineering decisions, cost teardowns, MCP-ecosystem analysis, and the occasional founder essay. New posts most weeks. RSS feed.
Automated scanners flag any tool that uses child_process.spawn as 'critical'. Here's the real threat model for MCP proxies — and what we do about each.
Read post →
Windsurf bundles its own MCP servers by default. Most users have no idea what's running. Here's how to find out — and what each one actually costs.
Read post →
Cloud FinOps took 5 years to mature. AI FinOps is repeating the same arc — but the unit of cost is now the MCP tool call, not the EC2 instance.
Read post →
Real numbers from 30 days of Cursor + MCP usage. Per-tool cost, per-session distribution, and the 3 tools you should optimize first.
Read post →
Anthropic shows you total usage; it doesn't show you which MCP tool spent it. Here's how to attribute every dollar to its tool, server, and project.
Read post →
Pick the right layer for your AI cost question. A practitioner's guide to four overlapping observability tools, with concrete examples of when each one wins.
Read post →
Cursor uses one location, Claude Desktop another, Windsurf hides theirs in ~/.codeium, VS Code splits user vs workspace, Claude Code uses project-relative configs. Here's the unified detector.
Read post →
A teardown of 30 days of Cursor + Claude Desktop usage, broken down per server, per tool, per project. The cost curve is steeper than you think.
Read post →
An honest writeup of building the first cost-observability tool for the Model Context Protocol — what worked, what surprised me, and the numbers nobody talks about.
Read post →
Subscribe to the RSS feed.
New posts every Wednesday. No newsletter, no popup — just plain RSS at /blog/rss.xml.