OperationKit is open source and free to install yourself. This is me doing the part that eats a month: loading your real context in, standing up your server, wiring your models, and writing your first agents. Four phases. Nothing in it is locked to me.
Everyone building agents right now is renting the same models from the same three labs. That makes the agent itself a commodity on a short clock. What does not commoditize is what the agent knows: your customers, your past decisions, the deals you have open, the people you owe, the way your business actually works. The agent is the cheap part. The context is the part nobody can copy off you.
The problem is that every tool in this category ships empty. Installing one is a single command and takes a minute. Getting it to the point where it knows your business and does work you would have paid a person to do is the part that takes weeks, and every product in the space leaves that entirely to you. That gap between "it is running" and "it is finished for my situation" is the real cost, and it is where people quit.
I already crossed it. My companies run on this: nineteen agents, more than eighty skills, and a markdown knowledge base I have been compounding for years, all on hardware I own. So I am not designing your setup from a blank page. I am porting a working one onto your box and filling it with your context instead of mine.
The knowledge base is plain markdown files on your disk. Every file is text with a small structured header on top, organized into folders per part of your business, readable in any editor and committable to git. I build the structure around how you actually work, then backfill it with the context you already have: your exports from Claude and ChatGPT, your email, your meeting transcripts, your notes, your contact records. Those get parsed into contacts and context documents so that on day one your agents already know your customers, your positioning, and your history instead of asking you to explain it every time. This is the phase that decides whether the whole thing is useful, which is why it goes first.
I stand OperationKit up on a server you own and pay for, either a fresh cloud instance or your own machine. It runs as a small set of containers: the application, a model-routing proxy, and its database. The installer generates every internal secret locally on your box, so the only things anyone types in are a model provider key and a domain name. Every service binds to the loopback interface and the only thing exposed to the internet is a single HTTPS reverse proxy with an automatic certificate. Your agent transcripts, your knowledge base, and your credentials sit on that disk, and I do not hold a copy.
Every model call routes through a proxy layer that sits between your agents and the providers. You put your own API keys in it, and they stay in the environment file on your server. Today that covers Anthropic, OpenAI, and Google, plus a local open-weights model running on your own GPU if you want work that never leaves the building. You configure named groups, so "the model that does research" and "the model that writes code" are settings rather than code, with an ordered fallback list underneath each one: if the first provider is rate limited or down, the next one takes the request. Changing providers later is an edit to one config file, not a rewrite of your agents.
The system has three layers with a one-way dependency: tools, then skills, then agents. A tool file documents one connection, meaning the endpoints, which credential reaches it, and the gotchas, with the names of secrets and never the values. A skill file is one repeatable job with a checkable finish condition. An agent file is a persona plus which skills it loads and when to hand off, written as a YAML header followed by a markdown body. I write your first three complete sets against real jobs in your business, and I write them in front of you so you can see how a request routes from an agent down to a skill down to a tool. There is a free catalog of eight first-party agents to start from, and every file I write for you is the same kind of markdown you can edit or throw out.
| knowledge_base/ | your markdown vault, structured and backfilled |
| server/ | a running OperationKit install on infrastructure you own |
| model_routing/ | your keys, your providers, named groups with fallbacks |
| agents/ | three agent files, written against real jobs |
| skills/ | three skill files, one repeatable job each |
| tools/ | three tool files, one connection each |
| runbook.md | how to restart it, back it up, and upgrade it |
| handover/ | a working session where I show you how to write the next one |
Everything on that list is a file or a machine in your name. There is no account with me, no license to renew, and no piece of it that stops working if you never speak to me again. OperationKit itself is Apache-2.0, so the software underneath is yours on the same terms whether I am involved or not.
It is a fixed fee for a fixed scope, quoted per engagement, invoiced against phases. I will not put a number on this page because the honest range depends on how much context you are handing over and how tangled your existing systems are, and any number I posted would be wrong for most of the people reading it. You will have a written scope and a fixed price before anything starts, and separately you will pay your own cloud provider for the server and your own model providers for the tokens. Those two bills go to you and I never touch them.
The four phases run in order because each one depends on the last, and the schedule is driven almost entirely by phase one. If your context is already in files, that phase is fast. If it is spread across five inboxes and a decade of documents, it is not. I scope the timeline in writing after we talk and before you commit to anything, so you get a real date rather than a range from a landing page.
All of it. The server is in your name on your cloud account, the files are on its disk, the API keys are yours, and OperationKit is Apache-2.0, which means you can run it, modify it, and use it commercially without asking anyone. There is no hosted component of mine in the middle, no license key that expires, and nothing that phones home to me. Fire me halfway through and you keep whatever phases are done and working.
You do not need to write code, and you should not need to touch a terminal to run the system day to day. You do need to be someone who is comfortable owning a server and an API key, and you need at least one person on your side who will open a markdown file and edit it after I am gone. If your honest answer is that nobody there will ever do that, see the section above about who this is not for.
Anthropic, OpenAI, and Google today, through one routing layer that all your agents talk to. You can also point it at an open-weights model running on your own GPU if there is work you want to keep entirely in the building. You are not choosing one, either: you name groups like "research" or "code," give each an ordered list of models, and the next one picks up the request when the first is rate limited or unavailable. Swapping a provider later means editing one config file.
It sits on your server. The knowledge base, the agent transcripts, and the encrypted credentials are all on that disk, the services bind to the loopback interface, and the only thing facing the internet is the HTTPS proxy in front of the app. The one place data leaves the box is the model call itself, which goes to whichever provider you chose, under your API key and your agreement with them. Route it to your local model and it does not leave at all. I do not keep a copy of anything and there is no account with me for it to live in.
Email me a few sentences: what you run, what you want the agents doing, and what your context is sitting in right now. If it is a fit I will send back a scope and a fixed price. If it is not I will tell you that, and usually point you at the free version, because installing OperationKit yourself costs nothing and the repository is public.
m@mikeluka.comI take a small number of these a year, so I answer every one myself.