Pay-per-token AI inference over Lightning. Non-custodial, open, censorship-resistant.
SAIL is a decentralized network where anyone can run an AI model behind a Lightning paywall and anyone can use it — discover a host, send a prompt, pay per output token, stream the response. No accounts, no API keys, no middleman holding your funds or your data.
- Non-custodial. You pay host invoices straight from your own Lightning wallet. The app never holds your funds.
- Open-core. The protocol and reference client are open (MIT) and forkable — that's the moat and the decentralization. Run your own host or client; nobody can lock you out.
- Censorship-resistant. Hosts are discovered over Nostr and reachable over Tor
(
.onion), so there's no central endpoint to block, and a host's IP stays hidden. - Pay for what you use. Metered per output token over L402; with an NWC wallet you pay per-chunk as it streams.
Grab the latest build from Releases:
- AppImage (portable, no install):
wget https://github.com/sailprotocol/sail/releases/latest/download/SAIL-x86_64.AppImage chmod +x SAIL-x86_64.AppImage ./SAIL-x86_64.AppImage
- Debian/Ubuntu (
.deb): downloadSAIL_*_amd64.debfrom the release andsudo apt install ./SAIL_*_amd64.deb.
The app bundles everything (Python runtime, Tor) — you don't need Python or a Lightning node installed to use it as a client.
Open SAIL, pick a host, type a prompt, hit Send, and pay. Two ways to pay:
- NWC (automatic, recommended). Paste a Nostr Wallet Connect string from a supporting wallet (Alby, Zeus, Coinos, Primal…). SAIL pays each chunk automatically as the response streams.
- BOLT11 (any wallet). No NWC? SAIL shows a single QR + copyable invoice for the request — pay it from any Lightning wallet (Strike, BlueWallet, Phoenix, Cash App…). Once it settles, the response streams.
Either way, the payment goes from your wallet directly to the host. SAIL never custodies funds.
Have a GPU? Serve a model and earn sats per token — no app to download to host. Hosting is a script, a browser wizard, and a dashboard:
- Download & run the install script.
scripts/install-host.shsets up the prerequisites (GPU driver check, Ollama, Tor), the repo + Python venv, and your model. It's meant to be read first, then run — not piped blindly into a shell. - Finish in your browser. The daemon opens a local setup wizard (detect → model → pricing → payout → seed backup → go-live). Pick phoenixd (recommended, self-custodial — the wizard installs and provisions it for you), or bring your own LND or an NWC wallet.
- Manage from the dashboard. A local web dashboard shows earnings/activity and a built-in wallet — balance, receive/fund, withdraw, close/sweep, and recovery-seed backup. The operator surface is localhost-only and never exposed over Tor.
Going live installs a systemd service, so the host auto-restarts and survives reboot, coming
back on the same .onion and pubkey.
📖 Run-a-host guide — step-by-step (fast path + manual). 💸 How payments work on your host — pay-to-open channels, fees, your recovery seed, withdraw, and close/sweep.
phoenixd hosts — the channel cliff. A fresh phoenixd node can't receive until an inbound channel opens, which needs a one-time inbound payment of ~25–35k sat. Until then the wizard and dashboard say "running, but can't receive yet" rather than "live to earn". Cross it by paying your node a one-time ~25–35k sat inbound from another wallet — details in the payments explainer.
- The host can see your prompt and response. Inference runs on the host's hardware in plaintext — this is not private/confidential inference. Don't send anything you wouldn't want the host operator to read. (A TEE-backed private tier is on the roadmap, not here yet.)
- Beta software. Expect rough edges. Linux x86_64 only for now (macOS/Windows later).
- Keep amounts small. It's early; use small sats while the network and clients harden.
- You are responsible for your wallet and your spend. Set a sensible max-tokens / wallet budget.
MIT — see LICENSE. © 2026 SAIL contributors.