A little recorder.
Real app context.

01 / A SMALL PACKAGE. A PRIVATE SERVICE.

One recording, end to end.

The SDK captures audio and explicit application context. Your Sup workspace stores the evidence, and a server-side Soniox connection supplies live transcription. No Soniox secret is shipped in your app.

02 / MEET SUPKIT

Add it at the root.

Add the Swift package to your Xcode project and link SupKit. Configure your registered public app ID and attach the overlay once. Use an authenticated enrollment backend, or expose manual pairing in your settings screen when that is the appropriate host flow.

import SupKit

Sup.shared.configure(.init(
    endpoint: URL(string: "https://voicesnap.xyz")!,
    appID: registeredPublicAppID,
    presentationMode: .recorderOnly
))

// Attach once; use your real privacy gate.
RootView()
    .supOverlay(isVisible: !isLocked)

// Optional manual pairing controls in Settings.
SupSettingsView()

registeredPublicAppID is the UUID from your workspace. RootView and isLocked represent your own app view and privacy state. These snippets show the integration shape; they are not a complete application.

Give the user a clear permission choice.

Add NSMicrophoneUsageDescription to the app’s Info.plist. Microphone access is requested on the first recording attempt. Your backend verifies the host account before enrollment; for manual pairing, the owner approves the device code in Projects. Approved credentials are stored in the device Keychain.

The overlay stays hidden for unenrolled devices and protected content. The SDK has no BabyDaze-specific dependencies.

03 / NAME THE MOMENTS THAT MATTER

Good context is intentional.

Use stable screen paths and control identifiers. sup captures a tagged target at touch-down, before its action changes the view. Areas without annotations remain explicitly unidentified.

SettingsView()
    .supScreen("settings/profile")

Button("Save", action: save)
    .supTarget(
        "profile.save", label: "Save profile"
    )

TextField("Email", text: $email)
    .supSensitive()

Sup.shared.state(
    "selected_tab", value: "timeline"
)

Keep annotations and breadcrumbs free of personal data, tokens, private URLs, and raw console output. Labels should describe the control, not the person using it. The SDK’s screen names come from your code; Soniox does not identify the interface.

04 / SHARE THE MICROPHONE

Play nicely with the host app.

A host camera can capture the microphone once and forward the same live PCM stream to sup. Feedback keeps recording without competing with the camera for the device input.

// While the camera owns microphone capture.
Sup.shared.beginHostAudioCapture()
Sup.shared.appendHostAudioPCM16(cameraPCM)

// After the camera releases the microphone.
Sup.shared.endHostAudioCapture()

// Finish feedback before a host account change.
Sup.shared.accountBoundary()

The forwarded stream is 16 kHz mono signed PCM16 little-endian audio. Apps that cannot forward their camera audio can still use the pause/resume handoff. A manual pause, phone interruption, privacy gate, or background transition requires an explicit resume. Reopening the app never silently starts the microphone.

For app-specific private or system interfaces, call setScreenshotsProtected(true) while they are visible. Check the package integration guide when coordinating audio playback and nested camera usage.

05 / PRESERVE THE ORIGINAL CLOCK

Timing is part of the evidence.

Microphone samples and UI events map onto one monotonic session clock. Transcript words use their audio timestamps, not their arrival time. Pauses remain visible, including after reconnects and delayed uploads.

• Audio and context are saved locally throughout the recording, targeting one-second persistence.
• Immutable segments retry safely. The service acknowledges them after durable storage.
• The server maintains the Soniox connection and persists provisional and finalized transcript updates independently of the phone.
• After an interruption, saved segments recover without restarting capture. An unfinished session is labeled interrupted; a connection loss alone does not prove a crash.
• Sessions expire after 30 days, with owner deletion available sooner.

Recording limits are 60 minutes per session, 60 minutes per workspace daily, and 180 minutes across the platform daily. Reservations are enforced across simultaneous sessions. Your workspace shows its current recording usage.

06 / FROM A RECORDING TO A CONVERSATION

“Read my latest feedback.”

The private inbox plays audio alongside its transcript and context. A workspace-scoped connection also exposes list_sessions, get_session, and search_feedback to Codex.

A companion skill guides evidence retrieval. You can ask what a tester meant, inspect the surrounding screen and tap events, and discuss a change. The connection is read-only; reviewing a recording does not authorize code changes or a release.

Authenticated access provides the data. The skill describes the workflow.

07 / THE SAME PROTOCOL ON THE WEB

The real little red button.

Import SupBrowser from /sdk/sup.js. Omitted presentationMode selects recorder-only; set 'companion' explicitly for Sup chat. A host supplies its own authenticated transport and project identity. Recorder-only mode lets users drag the dot without starting a recording.

Screenshot capture is optional. A host can pass screenshotProvider: createViewportCapture(), imported from /sdk/viewport-capture.js, for timestamped app-viewport stills with text entries and marked private content covered. This is not full screen-video capture. The Sup website itself does not enable screenshots.

Pending audio and opted-in images are stored in IndexedDB and removed locally after durable upload acknowledgement. Browser storage and suspension behavior require acceptance checks on target browsers. Owner login and device approval are required for this website’s recorder; opening the page does not start capture.

08 / TEST THE MOMENTS THAT ARE HARD TO FAKE

Before you ship.

The service and SDK have automated recovery and security coverage. Physical-device acceptance remains distinct from simulator compilation and synthetic imports.