LiveCap

Live captions and translation for everything you hear. Local-first, on your own AI.

Quick Start · How it Works · Bring Your Own AI · Privacy · Credits

---

Status: pre-MVP, building in public. The design package and full product spec are in this repo (docs/PROPOSAL.md, design/); the roadmap lives in the MVP epic. Star the repo to follow along.

What is LiveCap?

You're in a meeting, a webinar, or a call that isn't in your first language. LiveCap is a small pane of glass that floats beside it:

• It hears both sides — system audio (them) and your microphone (you) — and shows live captions in whatever language is being spoken, detected automatically.
• Under every sentence, it shows a translation into your language, half a beat later.
• A one-line live summary keeps you oriented; when the meeting ends, a clean Markdown transcript is already saved to a local folder.
• And when you need to speak up, reply suggestions and a quick-translate box give you the words.

Speech never leaves your machine: transcription is on-device Whisper. Translation runs on AI you already have — the Claude or Codex CLI installed on your Mac, or a bundled local model. No new subscription, no API key setup, no cloud account.

✦ Why LiveCap?

Every existing option makes you choose: cloud captioning services that stream your meetings to someone else's servers, OS captions locked to one platform and English, or meeting bots that visibly join your call as a participant. LiveCap is none of those — it's a private overlay only you can see (it's even excluded from screen sharing), it translates between real language pairs, and it works with any app that makes sound: Zoom, Meet, YouTube, a podcast, a person in the room.

─ How it Works

┌─ system audio ─┐                                    ┌─ caption (instant)
│                ├─► VAD ─► Whisper (on-device) ──────┤
└─ microphone ───┘              │                     └─ translation (~1s later)
                                │ finalized sentences
                                ▼
                   Translation engine (yours — see below)
                                │
                ├─ live summary · meeting board · reply suggestions
                └─ auto-saved Markdown transcript  (~/Documents/LiveCap)

Captions stream in word-by-word as speech is recognized. When a sentence completes, it's handed to the translation engine with rolling context and a glossary, so names and terms stay consistent across the meeting. The window comes in three sizes — a one-line capsule, a TV-style subtitle strip, and the full panel — and snaps to your screen edges.

─ The Engine is Yours

LiveCap doesn't ship a cloud backend. Translation, summaries, and suggestions run on one of three engines — your choice, switchable mid-meeting:

Engine	Cost to you	How
Claude / Codex CLI (auto-detected)	Currently covered by your existing Claude/Codex subscription	LiveCap drives the CLI you've already signed into, the way open-design does for design tools
Local model (bundled download)	Free	llama.cpp + a small instruct model, fully offline
Your own API key (planned)	Pay-per-use	For heavy users

CLI usage is covered by your subscription today. If Anthropic's policy changes so that programmatic use starts drawing on Agent SDK credits, LiveCap watches for it and auto-falls back to the free local model — captions never stop mid-meeting. An optional usage indicator in Settings shows where you'd stand if credits ever start to apply.

─ Quick Start

⏳ Not yet — the first runnable build lands with the MVP epic. When it does:

brew install --cask livecap        # planned

Open LiveCap → grant audio access → pick your language → it finds your Claude/Codex CLI (or offers the local model) → you're captioning. Three steps, under a minute.

─ Privacy

This is the part we take personally:

• Audio never leaves your machine. STT is on-device Whisper (Metal-accelerated). There is no LiveCap server, no telemetry, no account.
• Invisible to your audience. The overlay is excluded from screen capture — when you share your screen, others don't see your captions.
• Your transcripts are plain local files in ~/Documents/LiveCap. Grep them, back them up, point Obsidian at them, or turn auto-save off. Optional retention auto-deletes old ones.
• The only network traffic is the translation call to the engine you chose — and with the local model, there is none at all.

─ Roadmap

See the MVP epic for live progress. Headlines: macOS first (Apple Silicon), EN↔KO as the launch pair with auto-detection built in, Strip/Capsule/Panel modes, session archive, then Codex adapter, BYO key, and Windows.

─ Contributing

Contributions are welcome once the MVP scaffold lands. The short version of CONTRIBUTING.md:

• Branches: task/<issue-number>-<slug>, commits: [#<issue>] Short description
• No mock/temp/stub code in merged PRs — every PR ships the production implementation of its scope (the epic spells this out)
• Never commit credentials, transcripts, or model weights — .gitignore is set up to fight you if you try

─ Credits

LiveCap stands on excellent open source:

• Meetily (MIT) — the audio-capture + live-STT architecture this project builds on
• whisper.cpp (MIT) — on-device speech recognition
• open-design (Apache-2.0) — the local-CLI integration pattern
• llama.cpp (MIT) — local model inference
• Tauri (MIT) — the app shell

─ License

MIT — do what you want, just keep the notice.