Permissions, the change journal and its hash chain, encryption of stored secrets and protection from prompt injection are part of the core of Sedimenta — not a paid add-on. Your data stays on your server.
Sedimenta is software you install, not a service we operate. Nothing about your workspace is sent to us.
PostgreSQL on your own server, or a local database on one person’s computer that listens on 127.0.0.1 only.
Run fully offline: no calls to external services, with a screen that lists every outgoing request the installation could make.
Tokens and passwords of connected services are encrypted at rest with the installation’s own key.
A daily database and files backup out of the box. Keys are backed up only encrypted with your passphrase, and checked right after writing.
Export everything to readable markdown files at any time.
When someone leaves, wipe their personal data while their pages and records stay, signed “Deleted user N”.
Access is checked on the server for every read and write — in the interface, the API and over MCP alike.
Workspaces and spaces with roles, access inherited along folders, per-record access and explicit deny.
Each agent works with its own token and its own permissions. MCP clients such as Claude.ai and ChatGPT connect through OAuth; every connection can be found and revoked later.
Sign out on all devices in one step. A screen of permissions shows what every token is allowed to do.
A module writes only to its own records and tables, and only within the spaces where it is enabled.
A prompt can ask an agent to behave. Sedimenta makes the server refuse what breaks your rules, whoever sends it.
An agent reads the rules of a space before writing. A write that breaks a required rule is refused, from an agent or a person.
Where you want a person in the loop, agents write proposals with their source; a person accepts or rejects them.
Content an agent wrote after reading something from outside — a web page, an import, a connected source — is marked as untrusted, and dangerous actions are refused. Only a person can clear the mark.
Every change is recorded: who changed which field, when, and the value before and after.
Changes are chained by hashes, so the history can be checked for tampering, with a proof for any single change.
See what the system knew on any day in the past, and roll a change back.
Keep before/after values for as long as your policy allows. After that they are erased, while the fact of the change stays and the chain still verifies.
Write to security@sedimenta.ai. We reply within two working days and credit researchers who want it.
We’re working with a small number of teams that already use AI agents in their daily work.
Request early access