Insights
The Missing Layer
Introducing Introspective Context Engineering for MCP: a method for building Rich Domain MCP Servers where the AI doesn't just query your data, it understands what it means. Based on multiple MCP servers running on real production business data.
Read the Practitioner ReportWhat Is an MCP Server? Sixteen in Production for Agentic AI
An agent that can reach your data still does not understand it. What an MCP server is, and what sixteen of them in production taught me.
Don't Generate the UI. Let AI Compose It.
The tempting way to give an agent a real interface is to let it write the frontend. The better way is to give it a small set of trusted primitives and let it choose. A form, a chart, a map and a table, selected at runtime around the conversation.
Scale Capabilities Before You Scale Agents
Sixteen MCP servers, 103 tools, and no network of agents anywhere in the estate. Why the count that matters is capabilities, not agents.
We Replaced Jira With Markdown Files
One team, ten repositories, four tech stacks, and a Jira licence nobody enjoyed. We moved every ticket into markdown files next to the code. Seven months later, 15 projects and 12 people are on a board that costs nothing per seat.
The Ticket Is the Context Window
Once tickets were markdown in the repo, the format turned out to be the easy part. The real deliverable was the process, written down in a form the agent executes: one skill, two write paths, and a loop you can tell not to stop.
The Cheapest Time to Be Wrong
Every ticket passes three automatic review layers before a human reads a line of it: a second model on the plan, a multi-lens self-review before the PR, and an agent-driven bot review after. The earliest one is the one that matters.
Enterprise AI Without an Enterprise Budget
Nine production MCP servers in the first three months, sixteen by the end of September: no platform team, no framework dependencies, no API bill.
MCP Is the AI Platform
Sixteen production MCP servers and 103 tools at one mid-sized firm. No agent framework, no RAG pipeline, no AI-platform vendor.
Six Things the MCP Spec Should Fix
Six fixes the MCP spec needs after 90+ tools and eleven APIs in production: from Resources no client surfaces to enums that break agent reasoning.
Your MCP Server Should Get Smarter Every Week
A number that sometimes means hours and sometimes kilometres, summed into a confident wrong total. No confidence score catches that. Production telemetry does.
How to Write MCP Tool Descriptions: The 8-Block Pattern from 11 Production Servers
856 tools across 103 MCP servers: 97.1% have at least one description smell. Here's the eight-block pattern that fixed ours, from 90+ tools in production.
The Six Levels of MCP Servers
Not all MCP servers are the same. How the types differ, as a six-level ladder from hollow API wrappers to apps that write back. Drawn from eleven in production.
Your Data Is Fine. Your AI Doesn't Know What It Means.
Enterprise AI pilots fail 80% of the time. Surveys blame data quality. Across sixteen production MCP servers, the real problem was almost never the data.
Production MCP: A Practitioner's Guide
Nine production MCP servers in the first three months, sixteen by the end of September. The complete framework, from understanding data through identity-bound deployment.