Bot-Free Does Not Mean Local: The Three Privacy Architectures
As enterprise hosts and legal teams blocked uninvited recording bots from Zoom, Microsoft Teams, and Google Meet during 2025 and 2026, nearly every vendor added a "bot-free" desktop capture mode. Even historically bot-first platforms like Fathom, Otter, Fellow, Circleback, tl;dv, and Read AI now offer desktop audio capture.
However, for engineers, managers, and executives handling sensitive roadmaps, M&A, or personnel discussions, "no bot in the participant list" is not a security architecture. Every tool in the 2026 market falls into one of three concrete data-flow models:
- Model A: Cloud Stream (Granola, Jamie, Circleback, Fathom, Otter, Cluely, Read AI, Notta default): Captures system and mic audio on your Mac without a bot, but streams the audio to Deepgram, AssemblyAI, or AWS and sends transcripts to vendor-hosted LLMs. Granola and Jamie delete the audio immediately after transcription, meaning you can never replay the call or audit acoustic coaching metrics.
- Model B: Hybrid Local STT + Cloud LLM (MacWhisper Pro, Quill default, Hedy default, Krisp): Runs speech-to-text on Apple Silicon so raw audio stays on disk, but routes transcript text to cloud LLMs unless manually configured with Ollama.
- Model C: Gated Local-First (Metis Meet, Cued, Meetily, Anarlog): Video, dual-track audio, SpeechAnalyzer transcription, SQLite FTS5 + 512-dim NLContextualEmbedding retrieval, acoustic coaching metrics, and Apple FoundationModels run 100% on your Mac by default. When frontier cloud models or Cursor CLI are used, they require explicit per-Space opt-in and OS-enforced Seatbelt sandboxing.
