
Security dossier
How agent telemetry is isolated, encrypted, governed and investigated, with the limits stated as plainly as the strengths.
Control families / 7
TLS 1.3 in transit and AES-256 at rest.
Try it / runs in your browser
This is the same redaction rule set Tracify applies to inputs and outputs. Edit the text on the left. Nothing you type leaves this page.
Customer [REDACTED_EMAIL] called from [REDACTED_PHONE]. Card on file: [REDACTED_PHONE], SSN [REDACTED_PHONE]. Agent used [REDACTED_API_KEY] to refund.
Where a span goes
Four hops, each with a named job. The US queue hop is real, so it is on the diagram.
The SDK or OTLP exporter sends spans with a project key. Plaintext keys are shown once and stored hashed.
Email, phone, card, SSN and API-key patterns are removed before queued events leave the EU region.
Inngest Cloud processes queued span events in the United States, which is why we do not claim end-to-end EU residency.
Convex, Tinybird and Redis for the EU region are hosted in Ireland, scoped to the project that owns the key.
Plain limits
Storage is in Ireland, but queued span events pass through Inngest in the US. We say so instead of rounding up.
Runtime policies are observe-first. Cost ceilings are checked server-side and fail open, so monitoring never quietly breaks production traffic.
Project keys are shown once at creation. After that only the hash exists on our side.
Questions
EU region accounts, traces and cache are stored in Ireland. Queued span events pass through Inngest in the United States, so Tracify does not claim end-to-end EU residency.
Email, phone, card, SSN and API-key patterns are redacted before queued events leave the EU region.
Project keys are shown once at creation and only a hash is stored on our side.
No. Telemetry is scoped to the project derived from the authenticated key, and there is no cross-project query path.
Runtime policies are observe-first and cost ceilings fail open, so monitoring does not silently change or block production traffic.
Responsible disclosure goes straight to the people who can fix it. Affected users are notified within 72 hours.
