Security
A capability should only reach what it needs.
That principle shapes all of Kohentra’s security: your data stays in your workspace, every agent stays within its scope, and no capability ever grants itself the right to act.
Every company in its own workspace
Each organisation has its own data, settings and users, isolated from everyone else's. None can see another's data.
Trust is not shared either: an agent provided by Kohentra is the same everywhere, but what other workspaces have judged never counts in yours.
Read-only by default
Kohentra reads your Dolibarr, cross-checks, concludes and proposes. It changes nothing there until you have explicitly authorised a specific action. A decision taken on a situation never writes to Dolibarr.
Three locks before any write
Before executing, Kohentra re-reads the object in Dolibarr; afterwards, it checks the actual effect. A positive technical response is not enough to call it a success.
-
1. The platform
declares, for everyone, which actions can ever be executed. Today, just one: closing a ticket recognised as spam.
-
2. Your workspace
grants that exact capability, to this agent, in this version. Nothing by default. A new version starts with no rights.
-
3. The regime
propose, execute after confirmation, or without prior confirmation, never beyond the ceiling set.
A declared scope, enforced in code
Every agent version carries a visible security contract: what it reads, what may be passed to a language model, what it may request. Kohentra refuses everything else while the agent works. An instruction given to the model is never a security barrier.
Any widening (one more piece of data, data sent to the model, an action) must be explicitly confirmed, otherwise the new version is neither published nor activated.
Secrets out of reach
No agent can read an API key, a login or a token: no capability exposes them. No language model is ever given write access.
Everything logged, everything reversible
- Every agent conclusion is kept exactly as it was produced, with the evidence it had. Lost evidence is declared lost, never rebuilt.
- Every activation, rollback, permission or judgement is logged: who, when, on which version.
- Withdrawing a permission, putting an agent back into observation or rolling back to an earlier version is always possible.
- An execution reported as wrong suspends execution without confirmation until the incident is dealt with.
What we do not publish
We describe our principles, not our infrastructure: nothing that could widen an attack surface is published.
The site you are reading is independent of the application: it connects to no business database and to no Dolibarr instance.