Privacy

Waired is built so that your prompts and replies stay on your own devices.

Your data path

When one of your devices uses another’s model, the request travels straight between them over an end-to-end encrypted WireGuard link. It does not pass through any Waired-hosted service.

• The control plane only introduces your machines to each other. It distributes peer public keys and endpoints via a signed Network Map. It never receives your prompts or completions.
• The relay, used only when a direct connection isn’t possible, forwards encrypted WireGuard datagrams. It cannot decrypt them — it sees ciphertext, not content.

In short: the coordination service hands out keys; the conversation happens directly between your devices.

No silent fallback

Waired deliberately avoids “quietly send your data somewhere else” behavior. The one place a fallback exists — the Claude Code integration — is fail-open and visible: if your local serving is down, Claude Code falls back to the real Anthropic API so it keeps working, and you can see the routing state at any time with waired claude status or waired doctor. You stay in control of when your own model is used versus the cloud.

Cost and ownership

The model runs on hardware you already own, so there’s no per-message bill and no subscription. The Waired client is open source — you can read exactly what runs on your machines on GitHub. The coordination service that introduces your devices is the part hosted for you (the same split Tailscale uses).

Sharing controls

You decide which devices offer their engine to the rest of your network:

• waired inference share off keeps an engine private to its own machine while still letting you use it locally.
• waired pause takes a device out of routing entirely.

See Sharing vs. pausing.