Live comedy commentary for your Claude Code sessions — voiced by TV characters.
You're coding. Claude Code is doing its thing. But instead of watching a terminal scroll in silence, Gilfoyle and Dinesh are roasting your variable names. David Rose is having a breakdown about your merge conflicts. Moira is calling the codebase "bébé."
100% local. Your code never leaves your machine.
https://github.com/user-attachments/assets/demo-placeholder
The app watches your Claude Code session in real time, understands what's happening (user questions, code changes, build results), and generates character dialogue that narrates the session — synthesized with cloned voices and played through a floating Zoom-style widget.
Claude Code session (JSONL) → Event watcher (kqueue, <1ms)
→ Smart filter (keeps conversation, drops noise)
→ LLM generates dialogue (Qwen 3B local or Groq 70B cloud)
→ Voice synthesis (cloned voices via Pocket TTS)
→ Pipelined playback (synthesize N+1 while N plays)
The app never sits silent. Pre-written banter fills every gap. Voice-cached to disk — after first play, everything is instant.
The key innovation. Voice cloning normally takes 2-30 seconds per phrase. We eliminated that:
- Clone once — extract voice state from a 10s WAV sample, save as
.safetensors(~30s, one-time) - Load instantly — cached voice state loads in <100ms on every subsequent run
- Cache forever — every synthesized phrase is saved to disk as WAV. Same phrase = 0ms next time
- Pipeline — synthesize line N+1 while line N plays. No gaps between dialogue lines
After a few sessions, 90%+ of audio is pure disk reads. No compute. No latency.
This caching architecture is extracted as a standalone library: InstaClone
Drop a .claude/theater.md in your project and the app generates dialogue that knows your codebase:
# In any project, run:
/create-theaterThis generates a theme-specific config with:
- Project context — what the project does, tech stack, key concepts
- Custom few-shot examples — dialogue reacting to events that would actually happen in YOUR project
- Project-specific fillers — banter about your actual tools, frameworks, and pain points
- Term explainers — explains project-specific terms in character when detected
The app uses keyword-matched RAG to pick the most relevant example for each event batch. The 3B model doesn't need to figure out your project — theater.md tells it.
| Theme | Show | You play as | Vibe |
|---|---|---|---|
| Gilfoyle & Dinesh | Silicon Valley | Richard | Dry roasts, Big Head stories, Erlich callbacks |
| David & Moira | Schitt's Creek | Johnny | Dramatic overreactions, elevated vocabulary, bébé |
| Rick & Morty | Rick and Morty | Jerry | Interdimensional incidents, burp, existential debugging |
| Sherlock & Watson | Sherlock | Lestrade | Deductive reasoning, Watson translates to plain English |
| Chandler & Joey | Friends | Ross | "Could this BE any more...", food references |
| Dwight & Jim | The Office | Michael | Beet farm incidents, "FALSE.", looks at camera |
| Jesse & Walter | Breaking Bad | Hank | "Yeah science!", purity obsession, no half measures |
| Tony & JARVIS | Iron Man | Pepper | Stark quips, probability calculations, 3 AM builds |
| Professor & Bug | Original | The Student | Teaching mode — explains concepts for beginners |
You always appear in the widget as "You (CharacterName)". Each theme has full personality profiles, speech style rules, catchphrases, and incident-style few-shot examples. Characters tell stories ("remember when Big Head..."), not analogies ("it's like...").
- macOS 14+ (Apple Silicon)
- Ollama with Qwen 2.5:
ollama pull qwen2.5:3b - Python 3.10+ (for voice cloning sidecar)
git clone https://github.com/sameeeeeeep/SiliconValleyTheater.git
cd SiliconValleyTheater
make build
open build/SiliconValleyTheater.appcd TTSSidecar
python3 -m venv .venv && source .venv/bin/activate
pip install pocket-ttsDrop 10-15s WAV clips into Resources/Voices/<theme>/ (e.g., David.wav, Moira.wav). The app clones voices on first startup and caches them as .safetensors for instant loading thereafter.
Get a free API key at console.groq.com. Set it in Settings. Uses Llama 3.3 70B — much better dialogue quality than local 3B.
┌─────────────────────────────────────────────────────────┐
│ Claude Code Session (JSONL) │
└────────────────────────┬────────────────────────────────┘
│ kqueue (DispatchSource, <1ms)
┌────────▼────────┐
│ SessionWatcher │ tails latest JSONL file
└────────┬────────┘
│ AsyncStream<SessionEvent>
┌────────▼────────┐
│ EventSummarizer │ filters noise, keeps conversation
│ (no LLM) │ USER asks → CLAUDE responds → CHANGED files
└────────┬────────┘
│ buffered 30s
┌─────────────▼─────────────┐
│ TheaterEngine │
│ ┌─────────────────────┐ │
│ │ theater.md (RAG) │ │ keyword-matched project context
│ │ + few-shot example │ │ injected into every prompt
│ └─────────┬───────────┘ │
│ ┌─────────▼───────────┐ │
│ │ LLM Generation │ │ Qwen 3B (local) / Llama 70B (Groq)
│ │ 6 dialogue lines │ │ incident-style, in character
│ └─────────┬───────────┘ │
│ ┌─────────▼───────────┐ │
│ │ Voice Synthesis │ │ 3-tier cache: phrase → disk → live
│ │ (Pocket TTS) │ │ pipelined: N+1 while N plays
│ └─────────┬───────────┘ │
│ ┌─────────▼───────────┐ │
│ │ Gap Filling │ │ DynamicFillerPool: consume & replenish
│ │ (never silent) │ │ context-matched by event tags
│ └─────────────────────┘ │
└───────────────────────────┘
│
┌────────▼────────┐
│ Zoom Widget │ floating panel, always-on-top
│ + Subtitles │ speaker highlighting, sound bars
└─────────────────┘
| Phase | What Plays | Latency |
|---|---|---|
| App starts | Cold open intro (10 variations, consumed once each) | 0ms (voice cached) |
| LLM warming up | Context discussion about your project | 0ms (cached) |
| Events buffering | Filler banter (context-matched, from pool) | 0ms (cached) |
| Events ready | LLM-generated dialogue (6 lines) | ~2-4s first line, 0ms pipelined |
| Between batches | Term explainers or more fillers | 0ms (cached) |
| Idle | LLM generates new fillers in background | pre-cached for next gap |
From the menu bar icon:
- LLM: Ollama (local) or Groq (cloud, free, faster)
- TTS: Pocket TTS (local cloning), Kokoro (fast), Fish Audio (cloud), Cartesia (cloud)
- Theme: 9 character themes, assignable per room
- Buffer: How long to collect events before generating (5-120s)
- Volume: Master volume + per-session control
Sources/
├── Engine/
│ ├── TheaterEngine.swift # Central orchestrator
│ └── DynamicFillerPool.swift # Consume-and-replenish banter pool
├── Models/
│ ├── TheaterContext.swift # theater.md parser + RAG
│ ├── CharacterTheme.swift # 9 themes with personalities + examples
│ ├── FillerLibrary.swift # Pre-written fillers + term explainers
│ ├── Config.swift # Settings, per-session theme map
│ ├── DialogueLine.swift # Data models + dialogue parser
│ └── SessionEvent.swift # JSONL event parser
├── Services/
│ ├── DialogueGenerator.swift # Event summarizer + LLM prompt builder
│ ├── SessionWatcher.swift # kqueue file watcher for JSONL
│ ├── OllamaClient.swift # Local LLM client
│ ├── GroqClient.swift # Cloud LLM client
│ ├── TTSManager.swift # Pocket TTS sidecar manager
│ ├── FishAudioTTSManager.swift # Fish Audio voice cloning
│ ├── CloudTTSManager.swift # Cartesia cloud TTS
│ └── AudioPlayer.swift # Pipelined audio playback
├── Views/
│ ├── WidgetView.swift # Zoom-style floating widget
│ ├── VideoCallView.swift # Character tiles + user tile
│ ├── ContentView.swift # Full window with transcript
│ ├── CharacterPanelView.swift # Character tile with sound bars
│ ├── SubtitleOverlay.swift # Floating dialogue captions
│ ├── TranscriptView.swift # Scrollable history
│ └── SettingsView.swift # Configuration UI
└── SiliconValleyApp.swift # Menu bar app entry point
TTSSidecar/ # Python voice cloning sidecar
Resources/Voices/ # WAV samples per theme
MIT
Built with Swift, MLX, Ollama, and the belief that coding should be more entertaining.