DocsSign inInstall Kastel
Context files

When does a context file stop being enough, and how do you move on?

The threshold is the first document not everyone should see, and one is enough. Before it, a well-kept file is the right tool. After it, only two bad options remain: put everything in and expose it to every reader, or thin it down until it helps nobody. Moving on does not mean throwing the file away, since it becomes the first source of your context.

Published August 2, 2026

The file was the right answer to a real problem

An AI connected with no context knows nothing about your company. It invents plausible names, picks the wrong step in a process it has never seen, and states all of it in the same confident tone. Somebody at your end wrote the context file to fix exactly that, and they were right, because the answers improved the same day.

What the file repaired is the share of the knowledge everyone may read. The rest stayed outside, and for good reason: putting it in meant showing it to every person who opens a session. So the AI became competent on the general and stayed ignorant on the particular, which is precisely where you needed it.

The price of that gap does not appear on the model's invoice. It appears in the cost per useful result. An answer that misses gets corrected by hand, asked again, re-read, and an agent that gets it wrong three times costs three times, in the hours of people who already knew the answer.

The file is not thrown away, it becomes a source

Moving on is not a rewrite. What you wrote is company knowledge validated by somebody who knew it, which is exactly the material a structured context expects. The move happens in three steps, and the limits you ran into fall away one at a time.

01

Your file comes in as the first source

It is broken into items that each carry a subject, a date and an owner, instead of staying one block nobody re-reads. Other sources then join it, taken from where the knowledge actually lives, in file storage, in mailboxes or in the sales software. From that point, what changes at your end updates your context, rather than waiting for somebody to remember to reopen the document.

02

What you could never put in it can finally go in

The context knows who is reading it. The read scope follows each person's responsibility, and the AI working for them sees only what they are allowed to see. Salary bands, margin by account or a deal under negotiation can therefore join the context without becoming visible to everyone. You decide what each AI is allowed to see.

03

Every change becomes a dated decision

A correction no longer lands in silence. It is proposed, a human approves its exact effect, and the decision enters a sealed audit chain where each entry links to the one before. Six months later you can say what the context asserted, who changed it and who approved the change, which is the definition of a governed context.

Five moments where the difference shows

These five moments arrive at every company, and each can be replayed at yours in an evening with the free core installed. The last one does not favour us.

The momentWith a shared fileWith a governed context
A confidential document would help the AI a lotYou add it and every reader sees it, or you leave it out.You add it, and only the people whose responsibility it is can see it.
Somebody changes roleThe file does not know, and somebody has to remember.The scope follows the new role without anyone having to remember.
Somebody edits a lineThe line changes, with no recorded author and no approval.The change is proposed, approved by a human, then dated.
A client asks what the AI was able to readYou can show today's file, never the one from six months ago.The history of access decisions can be re-read, decision by decision.
You want to walk away with your workIt is a file. You already have it, and it fits in an attachment.Export is complete and available at any time, but you do have to run it.

The flagged line does not favour us. A file is trivially portable, and that is a real quality worth admitting.

Try it on your own file

Kastel's core is free, installs on your infrastructure with no size limit, and runs on your own AI keys. Take the file you maintain today and hand it over as the first source. Then create two people with different responsibilities and ask the same question as each of them.

The same file now produces two different answers, which is the point you were trying to establish. Then edit a line and look at what the chain records. None of this asks you to believe us, to talk to us, or to pay anything.

Below the threshold, keep your file

If you work alone, or as two or three people who tell each other everything, your file is the right tool and Kastel will not add much. The threshold is not a headcount, it is the first document not everyone should see. For as long as that document does not exist at your end, nothing this page describes is missing from your setup.

The move also has a cost nobody can pay for you. Deciding who is allowed to see what is organisational work before it is a matter of tooling, and it forces rulings on cases the file let you avoid. Kastel records and enforces the decision, it does not make it. The free core is also used from the command line and over MCP, with no web interface: the console comes with managed hosting or with the Enterprise plan.

See what your file becomes

The product page shows the full path, from your sources to the AIs you connect, by way of each person's scope and the record of every decision.

Explore the product