Trust & security

Recording and rendering are sandboxed

ffmpeg and the browser recorder run as separate operating-system processes in fresh namespaces with resource limits, a cleared environment and a single writable work folder. Rendering has no network access. Text never enters an encoder command line; it is written to files the encoder reads literally.

Hosted recording only reaches your allowlist

Every recipe has an origin allowlist. Hosted recordings resolve each allowed host, refuse private and reserved addresses, and pin Chrome to the checked address; every other hostname fails to resolve. Private and local apps are recorded by the companion on your own machine, which connects outbound and never opens a debugging port.

Secrets stay out of AI and out of recipes

Sign-in details are encrypted at rest and written only to a per-job file the recorder reads, then deleted. They are never sent to a language model, stored in a recipe or shown in a trace.

Claims are checked, not assumed

AI-written assertions are labelled “requires verification” until someone with the claim-approval permission links them to approved, fresh evidence. Final renders are blocked while a material claim is unverified.

No silent billing

Credits are held before work and consumed only on delivery. Failed renders release the hold, covered repairs never take credits, and exhausted allowances stop and ask instead of charging overage.

Roles and approvals

Owner, billing administrator, producer, editor, reviewer, client approver, analyst and read-only roles, with separate permissions for running screen automation, connecting accounts, approving claims, publishing, testing and authorizing spend. Every approval is recorded.