antiphon

antiphon.dev · web demo · docs · why this exists

---

Antiphon is a spatial-audio monitor for coding agents. Every agent session gets a voice and a seat in a virtual room around you: real binaural rendering (measured HRTFs, room acoustics), head-tracked by your webcam, gentle by design. Tool calls tick each agent's chord. Long tasks settle into a low machine hum. When an agent finishes, it tells you — a spoken sentence or two, from exactly where its work is.

And the part everything is built around: close your eyes. The camera notices, the desk falls away, and the room comes up. Turn toward a voice, linger, and it delivers its report. Open your eyes and a reply box is waiting.

Install

macOS 14+ on Apple Silicon:

1. Download Antiphon (Antiphon-<version>-macOS.zip), unzip, drag to /Applications, launch. Releases are signed and notarized.
2. Connect your agent — e.g. for Claude Code:

   claude plugin marketplace add cfoust/antiphon
   claude plugin install antiphon@antiphon
   

   Codex, OpenCode, Pi, and Aider adapters live in plugins/.

Wear headphones. That's the whole trick.

Voices default to the built-in macOS ones; add an ElevenLabs or OpenAI key in Settings → Voices for much better ones. Full instructions in the docs.

Try it without installing

The web demo runs the identical engine compiled to WebAssembly — same HRTFs, same room, head tracking through your browser — with scripted agents instead of yours.

Under the hood

One Rust DSP core (crates/antiphon-dsp), one C ABI, two hosts — the native app and the wasm demo are verified byte-identical (parity ≈ −155 dBFS, enforced in CI). Signal path: minimum-phase HRIR convolution + fractional-delay ITD → image-source early reflections → FDN or BRIR-convolution late reverb, all at 48 kHz in 128-frame blocks. A small Go daemon (antiphond, bundled in the app) registers agent sessions, assigns personas, and runs the TTS ladder. Details in the engine docs.

git clone https://github.com/cfoust/antiphon && cd antiphon
just bake   # bake the HRTF + room asset
just app    # build Antiphon.app (Rust staticlib + swiftc; no Xcode project)
just serve  # or: run the site + web demo locally

HRTF data: MIT KEMAR (Gardner & Martin, MIT Media Lab).

License

MIT © 2026 Caleb Foust