Every AI application eventually hits the same wall: the model is smart, but it is locked in a box. It cannot see your Jira board, your database, your files, your calendar. Before late 2024, every vendor solved this with bespoke plugins — one integration per app per data source, written against whatever API shape that app invented. M apps times N data sources meant M×N integrations, each one a snowflake. ...
Auth is the part of MCP everyone gets wrong first. I know because I build MCP gateways for a living, and the most common support question is some variant of “my client has a token, why won’t the server take it?” The answer is almost always the same: the token was minted for someone else. This post is a field guide. (If MCP itself is new to you — what servers expose, how the protocol talks, how it evolved — start with the companion post, How MCP Works, and come back.) Part one covers how each auth method works when a client talks to an MCP server directly. Part two covers what changes when a proxy — an MCP gateway — sits in the middle, which is where most enterprise deployments end up and where most of the interesting failure modes live. Every method gets a sequence diagram showing who talks to whom and which token moves where. ...
The most useful word I’ve added to my prompts this year is prove. When you ask a coding agent to “implement feature X and make the tests pass,” you get exactly what you asked for: passing tests. Often that means mocked dependencies, stubbed responses, and a green checkmark that proves the agent can write an assertion — not that the feature works. A mock is a mirror: the test passes because you told it what the answer should be. ...
In my last post I walked through five layers of token optimization that took a product classifier from $200+/month down to $25–40: context compression, two-stage prompting, exact-match lookup, similarity caching, and batching. Each layer attacked either the size of the context or the number of LLM calls. The post ended with the bill mostly tamed. But there was a layer I hadn’t built yet, and it’s the one that interests me most in hindsight, because it inverts the whole relationship: instead of making the LLM cheaper, you train a model to not need the LLM at all for most of the work — using the LLM’s own past output as training data. ...
A small piece of plastic in our kitchen soap dispenser snapped. The official Grohe replacement costs €80. I had a 3D printer and a Claude Cowork session — five iterations later, I had a working part.
The Problem One of the services I’ve been building for an ecommerce app is a product categorizer: given a product name, assign it a 3-level category path from a large taxonomy. The app is in Polish, so both the product names and the category tree are in Polish — which matters less than you’d think, since LLMs handle this well, but it does mean the examples in this post would normally look like “Drzwi sosnowe 80cm” instead of “Wooden door 80cm.” I’ve translated everything to English here for readability. Simple on paper, until you’re classifying ~1M products a month and watching your LLM bill climb past $200. ...
What the Hell Are Layer2 Rollups? If you’ve spent any time in the Ethereum or blockchain space, you’ve probably heard the term “Layer 2” thrown around—often in the same breath as “rollups.” But what does it actually mean? Why is everyone so excited about it? And why should anyone building on blockchain care? The Problem: Ethereum’s Scaling Bottleneck Ethereum is the world’s most popular smart contract platform, but it’s not without its flaws. The biggest? Scalability. The base layer (Layer 1) can only handle about 15-30 transactions per second: When demand spikes, fees skyrocket and the network slows to a crawl. Not exactly the “world computer” we all dream about. ...