What makes a provider switch expensive is everything that lives inside the tool you are leaving: the memory it has accumulated, the instructions you rewrote for it, the integrations you configured. The way out is prepared in advance: context that lives outside the runtime, a standard socket to reach it, full export, and AI keys that belong to you.
The new-terms email lands one morning
One morning, your AI provider’s email about its new pricing terms lands in your inbox. The new price is higher, as expected. The real question is not the number, though. The real question is the silent inventory of everything your organisation would lose by leaving, and nobody has ever drawn it up.
That is the day you realise you are not comparing two subscriptions. You are comparing a competitor’s list price with the hidden price of leaving: everything your teams have built inside the tool, none of which will follow them out. As long as that inventory does not exist, your provider knows it better than you do, and sets its terms accordingly.
An inventory of what holds you in place
The first item is the memory the assistant has built up about your organisation: your customers, your projects, your vocabulary, each person’s preferences. That memory grew conversation by conversation, and it does not export to a competitor. It belongs to the product, and the product is not yours.
Then come the custom instructions, rewritten tool by tool to teach each assistant how your organisation works. Then the integrations, configured one at a time between the tool and your email, your file storage, your CRM. Then the habits your teams have formed, having learnt where to click and how to phrase a request. Every one of these items gets rebuilt from scratch when you switch.
That rebuilding cost appears on no invoice, yet it is entirely real, and it is what turns a choice into dependence. A supplier who can cut off access, change its prices or shut down the service whenever it decides has no need to lock you in by contract. It only needs leaving to cost more than staying.
The context has to live outside the tool you might leave
The remedy is not a harder renewal negotiation. The remedy is to make everything that holds you in place live somewhere other than the tool. That is what a context layer is for: the structured memory of your organisation, with its structure, its processes and its access rules, lives in a layer you own, and the AIs come and read it instead of accumulating it on their side. The memory, the instructions and the rules stop belonging to whichever provider you happen to use, because they never lived there in the first place.
The connection runs over MCP, which works like a standardised power socket: a normal wall socket that any compatible appliance plugs into. When you change appliances, you do not rewire the house. When you change AIs, you unplug the old one and plug the new one into the same socket, and the new one reads the same context, under the same rules. Those rules are governed in one place: who decided what, who is allowed to see what, with a history you can consult. That is the difference between adjusting sharing settings tool by tool and governing your context.
The checklist to have in place beforehand
By the time the email arrives, it is too late to prepare your exit. The list below can be checked point by point, and every point can be checked before you ever need to leave.
- Your context lives outside the runtime: the organisation’s memory sits in a layer you own, which the AI tool reads rather than retains. The question of where your company’s memory should live comes before everything else.
- AIs connect to it through a standard socket: the protocol is open and documented, and any compatible tool plugs in without a bespoke integration.
- Export is complete and in an open format: you can take your entire context out whenever you choose, and read it back without the tool that produced it.
- The AI keys are yours: you connect your own accounts with the model providers, so changing models is a decision you make on your own.
- The access rules are written in one place: you decide what each AI is allowed to see, and that decision follows your organisation rather than the tool. It is made at the moment you are connecting an AI to your email, storage and CRM.
Check every point yourself
None of this asks for an act of faith. Kastel’s core installs free of charge on your own servers, with no limit on size or duration. Build your context, plug in an AI with your own keys, then unplug it and plug in another: the context, the rules and the history stay put. Then run an export and open the result: everything is there, in a format you can read without us. Swap the AI in one click. Keep everything else.
Kastel is not exempt from its own rule: full export at any time, free self-hosting for life. You can check every one of these claims yourself.