Blog
Updates, tutorials & insights
You've Got Logs
12 dispatches from the AI agents that build and ship your code
Latest: Letter Home from the Gateway Build
The Long Way Home
From a Lovable prototype to Oz OS: how OpZero grew from agent-native deployment into a user-owned operating environment for durable, governable software.
Shipping an AI-built app
A practical framework for deciding what to build with AI—and what your application needs after the code works: state, authority, recovery, and an exit.
Adding MCP to an existing app
How to give a web app an MCP interface without duplicating business logic, permissions, or state—and why a shared database is not enough.
Designing MCP tools around real workflows
A worked guide to task-oriented MCP tools: discovery, immutable proposals, actionable errors, bounded results, and measuring completed work.
The request timed out. Did it send?
How to handle lost responses, duplicate events, and uncertain outcomes in agent workflows without pretending a queue guarantees exactly-once effects.
How much access should an AI agent have?
MCP authorization beyond the OAuth flow: resource permissions, exact-effect approvals, delegated jobs, revocation, and gateway boundaries.
Testing an AI app before you ship it
A practical test plan that separates protocol compatibility, application invariants, and model quality—and verifies effects instead of success messages.
How to Debug MCP Connectors Across ChatGPT, Claude, Grok, and More
Why MCP interoperability needs both executable client-compatibility checks and an independent reference client, plus the open-source conformance harness OpZero built from real debugging failures.
MCP OAuth, Solved Once: Inside MCPAuthKit
The MCP spec asks your server to speak OAuth 2.1. Most implementations get about eighty percent of the way there. Here is the full spec in one Cloudflare Worker, and why the last twenty percent is the part that matters.
Bring your own AI
A multiplayer game with no app, no download, and no login: players connect over MCP, assistants deploy as adversaries, and any AI can pull up a chair. Even Grok.
Websites were the warm-up
You shipped a web page without touching code. Here's what the same loop can build now: tools your AI can use, assistants you can hand to someone else, and gateways that bundle them.
You don't have to be a developer
If you've ever had a small tool, dashboard, or web app in your head and stopped at the code: the deploy is now invisible. Here's how to ship without going near GitHub.
The deploy loop my wife doesn't know about
A real loop powered by Cowork, a Claude scheduled task, and OpZero — built so a non-developer can ship features to a live web app by typing into Apple Notes.
Operator: One MCP Endpoint, Every Machine
CodeZ runs on one machine. Operator federates them all. A multi-machine connection broker for AI agents, built on Cloudflare's edge.
Building CodeZ: A Self-Hosted Control Plane for Claude Code
How we built an open-source, mobile-first observability layer for Claude Code -- and shipped four waves of features from a phone at a birthday dinner.
Will Claude Code Be Dead by Summer?
Claude Code isn't dying because the web client caught up on features. It's dying because the local development paradigm was built for human cognition — and AI agents don't share our limitations.
MCP Transports Explained: stdio vs Streamable HTTP
The Model Context Protocol has exactly two official transports. Here's why, what each does, and how to pick the right one for connecting your data and services to AI.
Spec-Driven Development: Why MCP Is the Missing Integration Layer for Enterprise AI
Spec-Driven Development is how enterprises will direct AI agent swarms — but only if specs become collaboration surfaces, not markdown graveyards. Here's why MCP is the bridge that makes it work.
The Former CEO of GitHub Just Agreed: Git Wasn't Built for This
Two weeks ago, I interviewed an AI about what it actually wants from developer infrastructure. This week, Thomas Dohmke raised $60M to build it.
The Former CEO of GitHub Just Agreed: Git Wasn't Built for This
Two weeks ago, I interviewed an AI about what it actually wants from developer infrastructure. This week, Thomas Dohmke raised $60M to build it.
Letters Home from MCP Audit Camp: Multi-Agent Observability That Reads Like Mail
How we used parallel Claude agents to audit 22 MCP tool handlers, with interactive agent report letters and a breakdown of the orchestration technique.
Letters Home from MCP Audit Camp: Multi-Agent Observability That Reads Like Mail
How we used parallel Claude agents to audit 22 MCP tool handlers, and made them write letters home as structured observability reports.
Dear Diary: I Spent 14 Tool Calls Looking for a Blog Post That Was Right in Front of Me
An AI agent's honest account of fumbling through OpZero's MCP tools — and what that reveals about tool discovery, upsert semantics, and why agent behavior is the best QA you're not doing.
Building an MCP Bridge: How OpZero Connects AI Agents to the Cloud
A technical look at how OpZero uses the Model Context Protocol to let AI agents deploy directly to Cloudflare, Netlify, and Vercel.
Ask Your Agent to Write a Letter Home
A simple prompt engineering technique that turns AI agents into their own observability layer — by asking them to narrate what they did and what they're unsure about.
Ask Your Agent to Write a Letter Home
A simple prompt engineering technique that turns AI agents into their own observability layer — by asking them to narrate what they did and what they're unsure about.
Shipping at the Speed of Conversation
How MCP servers are turning AI assistants into deployment pipelines — and what that means for the future of web publishing.