Services / 01 / AI agents & MCP integrations
Give your AI access to the tools and data your business actually runs on.
I build MCP servers and agents that connect Claude, ChatGPT or your own model to your systems, with written rules for what the agent can read, what it can change, and what it has to ask about first.
Tell me about your projectFree 30-minute intro call. Written scope and quote before any work starts.
Sound familiar?
If any of this is you, keep reading.
Your team pastes the same data into ChatGPT every day, then pastes the answer back.
An agent demo worked in the meeting and fell over in its first week of real use.
The data you need sits in an app with no proper API.
Nobody has decided what the agent may do without a human signing off.
What you get
The deliverables.
- An MCP server or agent wired into your systems: CRM, database, documents, internal tools, or the app that has no API yet
- Permissions that are written down and enforced. Read-only by default, changes behind approval where it matters
- A log of every action the agent took, so you can check its work and fix what it gets wrong
- A model picked for the task and the bill, not the hype: Claude, GPT, Gemini, or one you host yourself
- Source code, documentation and a handover session
Proof
Things I've already shipped.
Senior developer for AI and blockchain at Angor Ag since 2024, shipping LLM tooling into production.
garmin-mcp
AI Tooling / Personal Data · 2026Garmin has no personal API, so this keeps a local SQLite archive of Garmin Connect data through one rate-limited ingest, then exposes it to Claude as an MCP server and writes it into Obsidian daily notes. When the unofficial endpoints break, only future data is lost, never the archive.
Morning Brief (Kindle)
E-Reader Tools / LLM · 2026KOReader plugin for a jailbroken Kindle. One tap pulls full-text AI and tech news plus Macedonian outlets into a single morning document, alongside a reading list, Reddit feeds, tasks synced both ways with Google Drive, and an LLM assistant.
Electric Cherry
Hardware / E-Bike Build · 2026Belt-drive e-bike conversion: a 6374 motor on a one-piece 3D-printed holder modelled in Fusion 360 and scripted over MCP, an HTD 5M belt at 7.67:1, and a Flipsky VESC on a 6S5P pack. Every fix, from idler creep to belt rub, gets measured and logged. The wireless link became its own open-source project: a €1–2 ESP-01 WiFi bridge for VESC Tool.
How it works
From first call to handover.
- STEP_01
Intro call
30 minutes. You explain the problem, I tell you honestly whether I'm the right person for it.
- STEP_02
Written scope
What gets built, what doesn't, how long it takes and what it costs. In writing, before any work starts.
- STEP_03
Build in the open
You see working progress every week instead of a big reveal at the end.
- STEP_04
Handover
Source, files, documentation and a walkthrough. You own all of it.
After it ships: Agents drift as models and your data change. After launch I can stay on for a few hours a month to read the logs, tune prompts and add the tools your team asks for next.
QUESTIONS
Which AI models do you work with?
Claude, GPT, Gemini and open models you run yourself. I pick per task, and I'll tell you when a cheaper model does the job just as well.
Does our data get sent to an AI provider?
Only what a task needs, and we agree on that before I write any code. If something can't leave your servers, we use a model that runs on them.
What is MCP?
Model Context Protocol, an open standard for connecting AI assistants to tools and data. One MCP server works in Claude, ChatGPT, Cursor and other clients, so you aren't locked into a single vendor.
Can you connect a system that has no API?
Often, yes. garmin-mcp is one example: Garmin offers no personal API, so it keeps a local archive and exposes that to Claude. We check the system's terms of use before building anything.