Comparison
Where captured screens, audio, and transcripts live, which models run on your laptop, and what actually leaves the device.
What is different in practice
With Overshow, OCR text, audio transcripts, and metadata sit in a local SQLCipher-encrypted SQLite database on the laptop. No screen images or video files are persisted. Cloud assistants upload the same content to their provider by default.
Overshow runs FluidAudio Parakeet TDT v3 for transcription, FluidAudio Sortformer fastV2_1 for speaker diarisation, EmbeddingGemma 300M for embeddings, and Gemma 4 E2B via MLX Swift for chat. All on-device. Cloud assistants call an external API for each of those.
By default, nothing does. The optional MCP server exposes approved tools to AI clients you choose, and only while the app is running. Most tools retrieve local context; a few update local metadata. Exports are manual.
Fixed 20-30 second capture windows with a 1-2 second overlap tail.
768-dimensional vectors indexed locally alongside FTS5.
Runs via MLX Swift on Apple Silicon. No external LLM calls.
Install Overshow on one laptop, use it for a week on your real work, and see what you miss from a cloud assistant and what you gain from keeping data local.
Comparison guide
This is a practical comparison, not an ideological one. Both models work. They differ in where captured data sits, which AI runs where, and what you need to explain to your security team.
Capture, transcription, search, and chat all run on the laptop.
Nothing in that list requires a network call to process your captured content.
The desktop app is local. The account does sit in the cloud, because billing, device registration, and SSO need to. What stays on the machine is the captured content itself.
Optional paths where data does leave:
Cloud-first assistants are usually the right choice when:
Local-first is usually the right choice when:
