← Back to plumb
The details

Models, languages & how it's built

Everything the homepage keeps light. Summaries are bring-your-own-key; recording and transcription stay on your Mac. Here's the full list — and what's under the hood.

Models & providers

Bring your own intelligence.

Summaries run on the provider and model you pick, with your own API key. Plumb doesn't sell or bundle AI credits — you pay your provider directly, only if you use the summary feature at all. Local transcription needs no key.

Anthropic Claude
  • Claude Haiku 4.5
    Fastest · most economical
  • Claude Sonnet 4.6 Default
    Balanced — recommended
  • Claude Opus 4.8
    Deepest reasoning
OpenAI ChatGPT
  • GPT-4o
    OpenAI flagship
  • GPT-4o mini
    Fast · economical
Google Gemini
  • Gemini 2.5 Pro
    Google's most capable
  • Gemini 2.5 Flash
    Fast · economical
Pick the depth at send time: Concise Balanced Detailed

Local transcription · Whisper

On-device

Choose a model size to balance speed and accuracy. Models are fetched once from the public whisper.cpp repository and cached locally; the live preview always uses Base.

  • Base ~150 MB Fastest, least accurate
  • SmallDefault ~470 MB Fast, good accuracy
  • Medium ~1.5 GB Slower, more accurate
  • Large v3 Turbo ~1.6 GB Slowest, most accurate

Optional cloud transcription

Off by default

Prefer higher accuracy and automatic per-speaker diarization? Plumb can send audio to Deepgram (nova-2) instead of transcribing locally. This is clearly worse for privacy — your audio, including your voice, leaves the device — so it's opt-in and plainly labelled.

26 languages + auto-detect

Transcription supports English, Spanish, French, German, and 22 more spoken languages.

Under the hood

Built like it means it.

Trait-based dependency injection: every capability — capture, transcribe, diarize, summarize, store, update — is a trait wired in one place, so each layer stays swappable. That's also why a fully-local summarizer is a supported future direction.

Language
Rust (2021) — a Cargo workspace of 11 library crates plus a thin app binary
UI
Dioxus 0.7 native desktop window, styled with Tailwind v4 + OKLCH design tokens
Transcription
whisper.cpp via whisper-rs, Metal-accelerated on macOS
Capture
Apple's ScreenCaptureKit for system audio; cpal for the microphone
Storage
SQLCipher-encrypted SQLite with forward-only migrations; secrets in the keychain
Networking
reqwest over rustls (no OpenSSL); local-only tracing, no remote sink
Updates
ed25519-signed manifest + SHA-256 integrity, verified before an in-place swap
cargo workspace
corecapturetranscribecloudsttdiarizekbredactsummarizeupdaterstorageuiapp

Release build is size-optimized (LTO, stripped) — the app bundle is ~10 MB.