Skip to content

The boundary is the laptop.

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. With Overshow, screen text, transcripts, and metadata sit in one SQLCipher-encrypted database on the laptop; cloud assistants upload the same content to their provider by default.

One machine, one boundary
Local capture index (default)
Captured screen text not selected for publication
Meeting transcripts not selected for publication
Embeddings and search
Chat and Ask
Crosses the network
Sign-in, billing, device registration, update and continuity checks
Approved record and source excerpts you choose to publish

What runs where in Overshow.

Capture, transcription, search, and chat all run on the laptop. Nothing in this list requires a network call to process your captured content.

CapabilityWhere it runs
Screen captureFocused-window text through the accessibility layer first; sparse OCR fallback, images discarded
AudioMicrophone and system audio, transcribed on-device over short rolling windows
Speaker labellingPer meeting, on-device; profiles linked locally
SearchFull-text plus on-device semantic embeddings
Chat and AskA local language model on Apple silicon

What leaves the device, and when.

The desktop app is local. Account operations and release checks use the network, but those routine requests do not carry captured content. Captured content leaves only through an explicit action: publishing sends the approved record and its approved source excerpts to the organisation repository; client access can be revoked; feedback diagnostics are reviewable before submission; an export is a plaintext copy you place and manage yourself; and optional URL enrichment fetches link metadata from the sites you captured.

PathWhat crosses the network
Account and updatesSign-in, billing, device registration, and release checks use the network; those routine requests do not carry captured content
Memory MCP (Pro)Off until you enable it and approve each client; a cloud client sends its query and returned snippets to its provider
Feedback diagnosticsOnly when you choose to attach and submit a reviewable diagnostic, which can include captured text
Organisation repositoryOnly when you choose Publish: the approved record and its approved source excerpts are stored for authorised organisation members; the local capture index stays on the laptop
ExportsYou run them manually when you want to

Which fit is yours.

Where cloud-first is a better fit

  • You want the most capable frontier models without buying Apple silicon
  • You do not capture screens or audio, only chat transcripts
  • Your security posture already permits sending work content to the provider

Where local-first is a better fit

  • You record screens or meeting audio and would rather not upload that material
  • You work across clients, vendors, or regulated environments and need the boundary to be the laptop
  • You want semantic search over your own history without an external index service

Questions to ask before deciding.

Whatever you choose, these four make the trade-offs concrete.

  1. Does the assistant capture your screen or audio? If yes, where is it stored?
  2. Can the assistant run transcription and embeddings without a cloud call?
  3. What does exporting or deleting everything look like?
  4. What does a security reviewer have to approve: an API or a database file?

Questions security reviewers ask.

Why does a local-first app need an account at all?

Sign-in, billing, device registration, and release checks cross the network, but those routine requests do not carry captured content. The local capture index is processed and stored on the laptop. Separately, a consultant can explicitly publish an approved record and its approved source excerpts to the organisation repository.

Does it work offline?

Capture, transcription, search, and Ask all run on-device, so core processing does not depend on a connection. You need to be online to sign in and manage billing, and Pro periodically refreshes its signed entitlement. If its offline grace expires, an account retaining an existing Free entitlement falls back to its historical recall window until a refresh succeeds.

Does it work with Zoom, Google Meet, and Teams?

Yes. Overshow recognises common meeting apps and browser calls and transcribes their audio locally once a meeting is detected. For an unrecognised surface, start recording manually. There is no bot or per-platform service integration.

What hardware does it need?

A Mac with Apple silicon, at least 16 GB RAM, and macOS 26 (Tahoe) or later. Running transcription, embeddings, and a language model locally is why the floor sits where it does.

What happens if the vendor shuts down?

This is the question this page exists for, so here is our own answer. Your database is already on the laptop and nothing you captured is deleted. Capture, transcription, search, and Ask run on-device, so the app keeps working without a server; only the recall window depends on a live subscription.

Compare it against your current assistant.

Book a walkthrough to compare the boundary against your current assistant and discuss access for an eligible Mac.

macOS 26 (Tahoe) or later, Apple silicon, and at least 16 GB RAM. Access is arranged directly.