Blog
Engineering notes
What I learn building MCP connectors, multi-agent systems and the infrastructure under them. Every post comes from a system that runs in production, with the real problems left in.
-
How to use AI in the classroom without losing the learning
How to use AI in the classroom so students still learn: attempt first, hints not answers, quizzes with devices closed, and fair assessment without detectors.
-
AI lesson planning: what the trials show teachers
AI lesson planning saves teachers real minutes, not hours. What the trials found, where checking time goes, and seven routines from Teaching in the AI Era.
-
MCP connectors are the App Store moment for AI
Why MCP connectors for Claude and other assistants look like the 2008 App Store: a standard socket, a directory, and builders shipping capabilities into an AI that millions already use. Lessons from shipping six connectors.
-
How to give a Grok Bot tasks when there is no API
xAI's Grok Bot has no API for sending messages. Here is a working pattern that lets Claude or any automation delegate tasks to Grok Bots through an inbox and outbox on their shared computer, with the exact standing instruction to paste.
-
Claude custom connector with Google sign-in: fixing redirect_uri_mismatch
Building a remote MCP server for claude.ai with Google OAuth and FastMCP? The redirect_uri_mismatch error almost always means the wrong callback was registered, or on the wrong Google client. The two redirect URIs explained, with working code and checks.
-
Claude, Grok and Codex on one build: one plans, one writes, one reviews
A real build where Claude designed and planned, the Grok CLI wrote the code test-first, and the Codex CLI reviewed every task. What the reviewer caught, the incident that changed my rules, and when this beats Claude Code vs Codex comparisons.
-
Disposable Proxmox VMs as a sandbox for Claude Code, Codex and Gemini
A sandbox for AI coding agents that is a whole virtual machine: cloned from a golden image in seconds, on a walled network behind an allow-list proxy, created by Claude through an MCP connector and destroyed automatically when its time runs out.
-
An SSH MCP server for AI agents: reverse tunnels and the three ways they break
Giving Claude a shell on a remote machine through an SSH MCP server and a reverse tunnel. Three failure modes that cost me real time: a tunnel that loops back to itself, host keys that change on every restart, and an OpenSSH penalty that locks the connector out.