Build your own context layer, or start from the free core?
Both are defensible, and the dividing line is not where people expect. Building your own stays sensible for as long as the context serves one team that may see everything. The cost arrives afterwards: keeping sources current, tying read access to each person's responsibility, and holding a record of those accesses that an outsider will accept. Kastel's core does those three things, free and on your own infrastructure.
The prototype works in a week, month six is a different job
You pulled a folder of documents together, put a search index over it, and wrote a small server that exposes the whole thing to your assistant. Within days the demo is convincing. It is convincing because the corpus is small, because you wrote it yourself, and because you are the only person reading it.
What the prototype has not met yet is time. Your company's knowledge lives in file storage, in mailboxes, in the sales software and in people's heads, and it moves every week. Nothing you built notices. Six months on, the index recites an organisation that no longer exists, and an AI never doubts a text you hand it.
Nor does anything record who may read what. The second reader, the one whose access differs from yours, turns your search tool into an authorisation system. That is a different job with a different set of mistakes, and its failures are quiet: nobody will tell you that an answer contained a line its recipient should never have seen.
What has to be maintained for it to stay true
Connecting is not the problem. MCP is a standard socket for AI, a universal port through which an AI tool reaches a company's context. It is an open standard, and anyone can write a server that publishes documents on it. If that is all you need, write it, spend an evening on it, and you will have been right to.
The work sits behind the socket. It comes down to three obligations, each of which is a standing commitment rather than a feature you ship once. They are what the product takes on, and they are in the free core.
Sources move, and the context has to move with them
A useful context is built from the tools where knowledge actually lives, not from a frozen copy. That means a connector per tool, the authentication that comes with it, and a repair every time a vendor changes an interface. At Kastel those connectors are part of the free core, all of them, with none held back for a paid plan.
Read access follows the person, not the tool
Ticking boxes tool by tool does not hold at the scale of an organisation, because a checkbox knows nothing about roles. At Kastel the read scope follows each person's responsibility, and it moves when they change jobs. An AI working for someone reads only what that person is allowed to read. You decide what each AI is allowed to see, and the rule that decides it is written down.
The access record has to survive being re-read
An application log proves very little, because whatever writes it can rewrite it. At Kastel every decision enters a sealed audit chain in which each entry is linked to the one before, so tampering with history after the fact breaks the seal and shows up on verification. That is the difference between a shared context and a governed one, and it is what somebody auditing you will ask for.
What you carry, what the free core carries already
Every line can be checked without talking to us, by installing the free core and looking. Two lines do not favour us and are here anyway, because you would find them out regardless and it is better to read them now.
The two flagged lines do not favour us. They say what you would establish for yourself the moment you install the product.
Install it before you decide
Kastel's core is free, installs on your own infrastructure with no size limit, and runs on your own AI keys. The test that settles your question takes three steps. Give two people different responsibilities, ask the same question as each of them, and compare what the AI was able to read. Then change an access and watch what the chain records. Then export your entire context and check that all of it comes back out.
Kastel is not exempt from its own rule: full export at any time, free self-hosting for life. Proof, not promise.
When you do not need Kastel
If you want to give context to your own editor, on your own machine, for your own use, you do not need Kastel. A folder of files and the server you will write tonight will do the job, and they will do it better, because they will fit you exactly. The threshold arrives with the second reader who does not share your access.
Past that threshold there is still a decision to make, and it is a grid rather than a sales argument. The free core is complete and has no size limit, but it is self-hosted and used from the command line and over MCP, which assumes somebody at your end is willing to operate a service. If nobody is, managed hosting starts at €39 a month and brings the web console. Either way, the free core remains the official trial path.
The shortest route is still to install it
Kastel's core installs free on your infrastructure, from the command line, with your own AI keys. One session will tell you what you would have had to build.
Install the free core