This page describes what the software does with data. It is written from the code, not from intent, and it is the reference the Terms and the Privacy Notice point at. Where a limit exists it is stated as a limit.
Your workspace's own records. Workflow definitions and every version of them, executions with their inputs and outputs, run logs, human-task approvals, connectors and the operations declared on them, inbound API configurations, webhook endpoints and their delivery log, API keys, the audit trail, and AI usage counters.
Accounts. Name, email address, and workspace memberships. Passwords are never stored by this platform: authentication is delegated to Keycloak, which stores a salted hash.
Credentials for your external systems. A connector holds the secret
it needs to reach your system — a password, a bearer token, an API key,
an SSH private key. These are encrypted at rest with AES-256-GCM using
the instance's SECRETS_KEY, are write-only through the API (they can
be replaced, never read back), and are decrypted only in memory, at the
moment an operation runs. They never appear in a workflow definition, an
execution payload, a run log, an AI prompt, or a connector revision.
One PostgreSQL database on the instance you or your operator runs. There is no separate analytics store, no third-party logging service, and no copy held by the software's authors.
Every tenant-scoped table carries an organization_id and is protected
by PostgreSQL row-level security, with the tenant set per transaction.
The application's database role does not have BYPASSRLS. Cross-tenant
read, insert, update and delete attempts are covered by tests that run
against a real database on every change.
Outbound traffic happens in four places, and nowhere else:
Outbound HTTP is default-deny against loopback, private, link-local, CGNAT and multicast addresses, enforced both at DNS-resolution time and as a pre-flight check on literal IPs.
Using "Edit with AI" on a workflow sends that workflow's definition, its declared inputs and outputs, the catalog of connector operations available to it, and your message. Using it on a connector sends the connector's name, protocol, base URL, authentication KIND, and the shape of its existing operations.
No credential is ever included in a prompt, which is the reason the platform removed inline HTTP calls from workflow definitions: a document that gets sent to a third party must not be able to carry your key.
The provider is whichever one the workspace configures. With your own key (BYOK), the request goes to that provider under your account and this platform meters nothing. With the instance's platform key, usage is metered in USD per workspace per month against a quota.
If you use no AI feature, no data leaves for an AI provider.
RUN_LOGS_RETENTION_MONTHS).
The product's chosen window is 6 months (decided 2026-08-20); the
variable is opt-in per deployment, so a deployment that has not set it
keeps them indefinitely. Executions themselves are not pruned by this —
only the line-by-line log of how they ran.Security-relevant actions — publishing a workflow, changing connector credentials, creating or revoking API keys, restoring a revision, managing members, wiping data — are recorded with who, what, which object, and when. The trail is append-only by construction: the application's database role has SELECT and INSERT policies and no others, so it cannot be rewritten even through an application bug.
The same applies to connector edit history.
These are stated because finding them out later is worse:
SECRETS_KEY rotation is not supported. Changing it orphans every
stored credential; they must be re-entered.GET /limits. They are not
restated here, so they cannot disagree.This software has none of its own. The parties who see your data are the ones your operator chooses: the hosting provider running the instance, the email provider configured for account mail, and — only if an AI feature is used — the configured AI provider. Your own external systems are not sub-processors; they are the point.