Give any terminal AI agent a voice to talk with you, and the discipline to be worth listening to.
Your agents speak their work, take turns through one microphone, and answer back to the exact agent that asked. The frame glowing around this page is the actual product.
macOS, Windows, and Linux. No account. No cloud. No cost.
The glow is the product.
A luminous frame wraps your screen. It turns blue when an agent is speaking and green when it is listening for you, so you always know the state of the room at a glance, without reading a line of text. Click any edge while it is green to send. The same sweep you see here is the live interface.
Plug it in. Your agent gains a voice, a conscience, and a ceiling.
The cooperative floor
Many agents, one set of speakers. Exactly one speaks at a time, you always win the floor instantly, and every spoken answer routes back by certificate to the agent that asked. They never talk over each other, and never over you.
Recover, build, break, fix, repeat
A self-adversarial loop. The agent attacks its own work from disk, breaks what it built, fixes what breaks, and refuses to call anything done from feeling. It proves the work before it ships.
The verified ceiling
One operating discipline that runs a capable model at its real ceiling: search wider, prove harder, never ship a claim above its evidence. Honest posture, no hype, every time.
Rigor you can check, not adjectives.
Download it, then just tell your agent.
One download carries the voice, the proving loop, and the operating discipline. It runs offline, on your hardware, for free, and your agents use it without ever touching the source.
The app is in final packaging and signing for macOS, Windows, and Linux. The download opens here shortly.
Direct download, no app store, no fees. Works with Claude Code and any terminal AI agent.