You plug in your sources
Mail, file storage, CRM, ticketing, calendars, e-signature, payroll: 55 connectors ship with the engine, and a folder you simply drop in works just as well.
The mechanism
The best AIs, fluent in your business.In a fortress that you own.
Kastel isn’t an AI. It’s the fortress that holds your company’s knowledge, and you decide what each AI is allowed to see.
+50 more
Your Kastel
ClaudeGPTMistralOpen model
Plugged into your Kastel, the AI answers with your data.
“What is our customer refund window, and who can make an exception?”
Nothing.
What your Kastel holds
Not a search engine over your documents: a structured representation of your company that any AI can read.
The organisation
The departments that actually exist, who does what, who reports to whom.
The processes
How the work really gets done at your company, not how a handbook says it should: the real procedure, exceptions included.
Who decides what
Kastel’s pivot. One matrix drives two things: who is allowed to see what, and who a request goes to for approval.
The entities
Clients, projects, suppliers, products. Linked to each other and never duplicated: a supplier record exists in exactly one place.
The memory
The decisions taken, their date, their reason, and what they replaced. Nothing is erased: what is no longer true is marked as such.
Underneath, it is versioned text files, and they remain the single source of truth. That is what makes full export possible: your content is locked into no proprietary format.
How it gets built
The engine reads what already exists
It derives a first org chart, the real departments, the entities that keep coming up. Before the first question is even asked.
A short interview
The organisation, the key processes, who decides what. The engine never asks again for something already present in your sources: that is exactly where the speed comes from, and it is mechanical, not magical.
You validate
The engine proposes an org chart and a responsibility matrix. You correct it, you settle the ambiguous cases, you approve. Nothing is written without that approval.
What the engine needs to knowExample
Sources+50 more
How many departments?6 → 5correctedDrive · org-chart.xlsx
Who runs sales?Camille FabreGmail · 412 threads
How many active customers?214HubSpot
What is the customer refund window?14 calendar daysDrive · terms.pdf
How is a leave request approved?Manager, then HRNotion · internal guide
+ 9 more questions answered by your sources
Who can make an exception to the 14-day window?Head of Sales, above €500your callyou
Which decisions go up to the leadership team?Any commitment above €20kyour callyou
Who is allowed to see HR files?HR and leadershipyour callyou
17 questions to represent your company
Your Kastel takes shape in one session.
With no data team and no weeks of mapping. The actual duration depends heavily on the size and complexity of your company: it is what the product aims for, not an average measured across dozens of deployments.
How AIs plug into it
Kastel runs no model. It publishes your context on a standard socket, and the AIs come and work on it.
Quotes agent only reads Sales and Operations.
MCP, the standard socket
MCP is a standardised socket for AI, the universal port an AI tool uses to reach a company’s context. Kastel is an MCP server: it publishes your context on that socket, and switching AI means unplugging one appliance and plugging in another on the same wall socket, nothing to move.
Your own AI keys
Inference runs on your keys, straight at the provider you already use, or on a model running at your place. Kastel never resells inference and takes no margin on it: what a clean context saves goes into your pocket, not ours.
How your Kastel grows
AI proposes, your team decides.
Refunds
1
Refund window: 14 calendar days.
2
Above €500, the exception goes through the Sales Director.
audit chain
Kastel in action
A 120-person industrial company in precision engineering.
It connects an AI key and the sources already in place: the mailboxes, the file store, a few exports from the management system.
The owner
Sits through the scoping interview. The engine has already pre-filled whatever the documents said and proposes an org chart and a responsibility matrix. Approves, corrects, settles the ambiguous cases.
The methods manager
A month later, plugs in the AI of their choice and asks: “what is our qualification procedure for a new supplier on a critical part?” The AI answers correctly, because it reads the company’s real process, and only what its scope allows it to see.
The quality manager
Approves the changes an agent proposes on a supplier record, down to the exact effect. Everything is recorded.
Nobody had to learn a concept. The owner approved an org chart, the quality manager approved a change.
Within two to three years, all AI models will be roughly equivalent. The advantage will no longer come from the model, that layer is commoditising, but from having structured and owning your company’s context.
And it belongs to you
Only now, the question of ownership. It is proven by facts, not by a word.
Your data, your context, your infrastructure
Kastel runs on your own server if you want it to, and the complete core is free there, forever.
Full export
One command, and everything comes out: context, governance, audit logs. No proprietary format to crack, they are text files.
Open interface
The protocol AIs plug into, the export format and the connector contract are open and standard. The engine stays our recipe: you own your data and your context, not our code.
Hostable in Europe
If you would rather operate nothing, we host and run an isolated instance for you, in Europe. Never one engine shared between several companies.
Kastel does not exempt itself from its own rule: full export at any time, free self-host for life. Proof, not promise.
Let’s build your fortress.
The complete core is free to self-host. If you would rather we host it, or help you build it, tell us.
Would you rather have help? Our partner integrators take care of it.