Clarity

Clarity, a voice that guides.

A macOS voice agent that watches your feed with you and speaks up the moment something's off — detecting emotional manipulation, flagging misleading framing, and guiding you back to clarity in real time.

Powered by

What it is

An always-on voice agent that turns doom-scrolling into informed watching. It reads the emotional manipulation in content and tells you what's really going on — out loud, the moment it happens.

It doesn't summarize after the fact. It intervenes live, with a calm voice, the instant your feed tries to play you.

How it works

Built on ElevenLabs for low-latency conversational voice that speaks at the exact moment you need guidance — Scribe v2 Realtime transcribes your feed as it plays, and an ElevenLabs voice cuts in to name the manipulation move. Claude Sonnet 4.6 does the reasoning that decides when and what to say; Tavily verifies the factual claims before the voice calls them out.

Not a logo flash — the agent is the product.

system audio ─▶ ScreenCaptureKit (Swift) ─▶ local WebSocket ─▶ Python server
     ▲                                                              │
     │                                            ElevenLabs Scribe v2 (live STT)
     │                                                              │
     │                                            Claude Sonnet 4.6 (classify the move)
     │                                                              │
     │                                            Tavily (verify factual claims)
     │                                                              │
  the voice ◀──────────────── ElevenLabs TTS (warm, decisive) ◀────┘

Why it's special

It doesn't scold you — it points at the content. The originality isn't fact-checking; it's the warm-voice emotional reframe: a voice on your side that names the trick being played on you, in the moment, so you can keep your feet.

What it catches

	Move it names
False	Misleading stats and fabricated claims — verified, then flipped
Toxic	Ragebait and outrage farming dressed as opinion
Empty	AI-slop and engagement filler with no real value
Empathy	Sad-bait and doomposting that feeds on loneliness
Medical	Health myths — redirects to professionals, never diagnoses
Safe	Stays quiet. Silence is the default.

Run it

Two terminals. Requires macOS 13+ and the keys below in a gitignored .env.

# 1 — Python server (uv, not pip)
uv run --env-file .env server.py        # listening on ws://localhost:8765

# 2 — Capture + menu bar app
cd Capture && swift run                  # grant Screen Recording permission, run again

.env needs ELEVENLABS_API_KEY, ANTHROPIC_API_KEY, TAVILY_API_KEY. Optional VOICE_ID.

Voice is off by default — flip it on from the 👂 menu bar item when you start scrolling.

See clearly. In a feed built on lies.