Serhii Kravchenko, MCP engineer
Yes MCP is my practice, not a company with a floor of engineers. One person does the work. That is worth saying plainly, because it also means everything published here comes from something I ran myself.
What I have actually built
A Model Context Protocol server that runs in production and gets used every day. Not a demo repository. It is a remote connector with proper authorisation: an OAuth 2.1 resource server, the RFC 9728 discovery document, tokens from a real identity provider, state kept per user, an audit trail, session limits, an idle timeout. A live user reaches it through Claude on the web, on a phone and on the desktop.
On top of it there are interactive interfaces, the ones the spec calls MCP Apps views. They render inside the assistant with real data. One is a card the user ticks off during the week, another shows the exact passage a quote came from. Both are shipped and checked on Claude's web and mobile clients. That is a narrower claim than it sounds, and deliberately so: most of what exists in this space has only ever been seen working in one client, on a developer's machine.
Then there are the measurements. Everything in the writing came out of that same server. Which channels between a server and a model actually arrive. What the protocol overhead really costs on every call. How one host behaves in a chat and differently in another mode. None of that came from reading the documentation, because the documentation kept turning out to be wrong. It came from cheap tests that could only answer yes or no.
The work that led here
MCP is the newest layer of something I have been doing for a while: putting AI agents inside organisations where people actually work. I was brought in on contract as Head of AI at Keabank, a fintech, where I built prototypes and took AI from an idea in a meeting to something teams used. Elsewhere I have brought AI into teams and trained the people who then had to live with it, and built agent solutions for specific departments, marketing, lead generation and IT among them. The recurring lesson is the one that shows up again in the MCP work: the model is rarely the hard part, and the boundary of what it may do on its own is where projects succeed or fail.
I also run a small private practice onboarding C-level executives onto agent systems. Most of them arrive having only ever used a chat window, and the useful shift is not learning prompts. It is seeing what changes when the thing can take actions in their own systems, and deciding where a human still has to sign.
Alongside that I am co-founder and head of R&D at a four-person startup, currently preparing a first investment round while testing in production. It builds AI systems for content and site operations. That work is not the subject of this site and I do not discuss its internals here.
What I do not claim
No MCP client cases yet. The agent and automation work above is real and shipped, but Model Context Protocol specifically is where my own production server is the evidence rather than someone else's project. When there are client cases in MCP, they will appear with the client's permission and with numbers that can be checked. Nothing on this site is invented.
I am also not a twenty-year systems engineer, and pretending otherwise would fall apart in the first technical conversation. I came to this from outside. Two years ago I left e-commerce with no technical background and went all in on AI, building real things until they held together. What that path left me with is a refusal to trust anything until it has been tested on a live system. In this technology that habit pays, because the specifications and the hosts disagree with each other constantly.
How these notes are produced
The findings are mine. They were measured on systems I run, and the commands and their output are kept. The prose is written with AI assistance, which seems worth saying out loud on a site about AI tooling. Every number here came from a command whose output I read. Nothing is estimated and then presented as measured. Where something is still a guess, it says so.
If something here turns out to be wrong, I would rather hear it and correct it in place with a dated note than quietly edit the page. Corrections go to the address below.
Elsewhere
I write about the wider path at AWRSHIFT: agent systems, AI search, and building from zero. That one is a personal blog, not a company. I am also co-founder of a venture in content and site operations that has not been named publicly yet. It is not what this site is about.
LinkedIn · GitHub · X · [email protected]
Who is responsible for this site
Yes MCP is operated by Serhii Kravchenko. Contact for anything, including corrections and takedown requests: [email protected].