Skip to main content

Tag · 16 essays

Cloud & infra

Writing tagged "Cloud & infra".

← All writing

Claude Desktop MCP lifecycle is broken

Claude Desktop currently launches duplicate local MCP stdio servers for a single configuration, causing two independent OAuth flows, extra browser tabs, and unnecessary resource usage. The bug is masked once credentials are cached, so it silently persists in production while still spawning two processes every time. This is not an mcp-remote or external service issue but a host-level lifecycle problem, likely caused by overlapping legacy and new MCP managers and poor observability for one of the processes. Workarounds like pre-authenticating with mcp-remote mitigate UX pain but do not fix the duplication. The article argues that open-source tools should not shoulder complex coordination logic just to survive a major desktop app’s sloppy process management.

  • Composable architecture
  • AI engineering
  • Performance
  • Cloud & infra

Ten AI security problems hiding in plain text

This article argues that AI security risks extend far beyond model jailbreaks and prompt engineering, into every piece of text an AI can read. Natural language now behaves like soft code, where logs, documentation, commit messages, wikis, support tickets, web pages, PDFs, and search results can all carry hidden instructions for agents. The author walks through ten concrete scenarios where ordinary text fields become attack surfaces, often with delayed or indirect activation through internal tools and multi agent pipelines. The core message is that information and instruction have blurred, and any text accessible to AI must be treated as part of the security model. Teams need to rethink access, editing rights, retrieval, and agent capabilities accordingly.

  • Composable architecture
  • AI engineering
  • Cloud & infra
  • Frontend

We are thinking too small

AI coding agents are not a threat to developer jobs, they are a fundamental shift in the economics of software creation. Just as the cloud removed the risk and capital cost of infrastructure, AI is removing the cost of writing and refactoring code. This kills the old moats that protected horizontal enterprise platforms and makes rebuilding wide, deeply integrated stacks viable. Instead of spending years stitching together narrow SaaS tools and APIs, teams can let agents generate bespoke services quickly and cheaply. The real risk is using AI only to speed up legacy glue work. Developers who win will treat the cost of reinventing the wheel as effectively zero and pursue much larger, previously impossible product ideas.

  • Composable architecture
  • AI engineering
  • Cloud & infra
  • Frontend

The case for boring setups

Most developers overconfigure their machines and pay for it in ways they don't notice. Every custom alias, every remapped key, every hand-rolled config file is a small tax on your ability to work anywhere other than your own laptop. The same pattern is repeating right now with AI tooling, dozens of MCP servers and stacked skill files that look like leverage and behave like drag. After getting stranded by my own setups more times than I'd like to admit, I've come around to a quieter belief: portability is a skill, and defaults are how you practice it.

  • Composable architecture
  • AI engineering
  • Cloud & infra
  • Frontend

I will not log into your vibe coded app

Vibe coding has made it trivially easy to spin up apps that collect user data, handle logins, and store sensitive information. The problem is that most people building these apps have no idea how authentication, database security, or data protection actually work. Defaults on services like Supabase and Firebase are dangerously permissive, GDPR obligations are being ignored entirely, and users are handing credentials to apps built by people who have never heard of row level security. My personal rule now is simple. If it looks vibe coded and asks for username and password, I walk away.

  • Composable architecture
  • AI engineering
  • Cloud & infra
  • Frontend

MCP fragmentation, context efficiency, and the rise of curated skills

The Model Context Protocol (MCP) was supposed to be a universal way to connect AI models to tools, but in practice it is fragmenting fast across vendors and implementations. Tool catalogs are extremely context-hungry, making naive MCP setups expensive, slow, and unreliable, especially with cheaper models. Developers are compensating with application-layer tricks like curated tool subsets, OAuth-based selection, Claude Skills style abstractions, and wrapping deterministic automation platforms such as Contentstack Automate. These patterns improve cost, reliability, and debuggability but highlight protocol-level gaps in context efficiency, determinism, and interoperability. The ecosystem is replaying past standards wars, and no clear winner is visible yet. The pragmatic move is to design flexible systems that can adapt when consolidation and better standards eventually emerge.

  • Composable architecture
  • AI engineering
  • API design
  • Performance