Before your IT team connects demi to scripts, systems or AI clients, they want to know: who may do what, where do the keys live, what is logged, and where does the data go? Here are the answers.
Permissions: never more than the person
- A key is a person. Every API key acts with exactly the permissions of the person who created it. Its permission levels (read, write, run agents) can narrow it further, never widen it.
- An AI client is a person. Over MCP every person signs in themselves. The client sees only what they may see.
- Agents get short-lived tokens. They act with delegated tokens bound to task, person and tenant.
- Checked on the server. Permissions follow from tenant, organisation, org unit, team, project and document tree. demi checks them on the server for every request, not in the AI model.
Manage keys safely
- API keys are shown exactly once. demi stores them only as hashes.
- Every key has an expiry: never, 30, 90 or 365 days. For scripts we recommend 90 days.
- You can revoke a key at any time in the app. A revoked or expired key gets
401. - Keys start with
demi_pat_. That lets secret scanners find them in repositories and logs. - In the CLI,
--api-key-stdinreads the key from a pipe. It lands neither in your shell history nor in the process list. - The config files for AI clients (
.mcp.json,.vscode/mcp.json) hold no secret.
Changes by AI clients
- Classification
- Every MCP tool is classified as read-only, changing or destructive.
- Sign-off
- Changing tools need a confirmation that is valid for that one action only.
- Blocked
- Deleting, ordering lead reports and creating API keys are blocked for AI clients.
- Evidence
- Every refusal is logged.
Logging
Every signed-in API request is logged with person or system, tenant, time and result, when it arrives and when it ends.
Tenants and encryption
- Each customer sits in its own tenant. The database also enforces the separation with row-level security.
- Data is stored with AES-256 and travels with TLS 1.2 or newer. Backups are encrypted too.
- Sign-in runs centrally over OpenID Connect, with single sign-on through your identity provider and an optional second factor by TOTP app or passkey.
Where the data goes
- Platform: demi runs in EU data centres, with Germany as the primary location.
- AI models: through EU endpoints by default. Other regions only if you agree them with us in your contract. More under Models.
- External AI clients: each person connects them with their own demi account and permissions. The answers of demi's tools go to the client's provider, because its model processes them. A switch for the whole tenant comes next.
- Training: demi trains no AI models on your data, unless you explicitly want it to.
Checklist for your IT team
Choose AI clients on purpose
Decide which external AI clients are fine for which data, and tell your team before they connect them.
One key per purpose
One key per script or system, with the shortest fitting expiry and only the permissions it needs. Keep it in a secret manager or your CI's secrets.
Open the network
Scripts and the CLI need HTTPS to
api.godemi.com, AI clients tomcp.godemi.com, the install todownload.godemi.com, plus the demi sign-in service.demi doctorchecks proxy and firewall; we send you the full list for your firewall.
Found a vulnerability?
Report it to security@godemi.com. How we handle reports is in our disclosure policy.
