Full settings for a model: install dtype + current runtime prefs.

* Get current streams

* Get all available capture targets (monitors and windows). * Desktop only.

* Get a token using the registered provider.

Get accumulated transcript since last clear (committed + interim)

Get all available capture targets (monitors and windows)

Get all available capture targets

Fetches user email from Auth0 by user ID. Uses in-memory cache to reduce API calls. Args: user_id: Auth0 user ID Returns:

Get the local IP address for network access.

Get the path to the loaded mmproj

Get the path to the loaded model

Settings.json management

Dependency to load and validate Twilio credentials.

Get whether GPU acceleration is enabled

Fetches a user's app_metadata from Auth0. Args: user_id: Auth0 user ID Returns: app_metadata dict or empty dict if not

Get valid access token, refreshing if needed

Give `email` a seat on `org_id`. If they already have an Observer account, entitle them immediately. If not, park a single-use invite in

Initializes the PiP plugin

* Check if buffer is empty

* Check if buffer is full (reached max duration)

* Check if a local model is ready for inference.

Check if tokens are expired

Check if multimodal (vision) is available

JPEG encode quality, 1–100.

The daily limit that applies to this user's tier for this service.

* Projector files (mmproj) that aren't paired with any model. * Used to populate the "assign projector" dropdown. * Orphans don't appear as thei

* Reusable utility to load an image from a base64 string. * Handles the data URI prefix automatically.

Logout - delete stored tokens

Max output width in pixels; the source is downscaled to fit (aspect preserved).

Stop capture if neither video nor audio is needed

Extract model ID from a filename (removes .gguf extension)

Normalize phone number to E.164 format using phonenumbers library. Handles all country codes and formatting variations automatically. Ex

Parse a target ID into its components

Fuzzy match two phone numbers using subsequence matching. Rules: 1. Both must have at least 8 digits (prevents "911" matching long numbe

* Push the persisted capture quality (max width / JPEG quality / FPS) to the Rust * backend. Called right before each desktop capture starts so the

Read new frame from the buffer Returns (jpeg_bytes, width, height, timestamp) if new data is available

Save base64 image to temp storage with compression and return public URL

Save base64 video to temp storage and return public URL. Args: video_b64: Base64-encoded video data max_size_mb: Maximum allowed

* Send a prompt through the appropriate backend for the given model. * - Local models (browser_local, llama_cpp_local): routed directly, no inferen

Store a new capture config. Values are clamped to sane ranges so a stray input field can't hand the capture pipeline a zero width or a 1000fps interva

* Set PCM callback for unified transcription pipeline. * When set, decoded PCM samples (Float32Array @ 16kHz) are passed to this callback * inst

* Set whether to use GPU acceleration (Metal) * Must be called BEFORE loading a model to take effect. * Setting is persisted to localStorage.

Store the audio channel from Tauri command

Set the audio channel for streaming

Store the video channel from Tauri command

Set the frame channel for streaming

Register a function that forwards log lines to the frontend. Call this once during Tauri setup before any engine activity. `f(level, message)` — level

Configures logging to output structured JSON. This is called once from api.py at import time.

* Start capturing system audio. * Audio is passed to Rust via JNI for resampling and channel delivery. */