SGL Node

Turn your Mac into a compute node on the Singularity grid — earn $SGL by serving confidential AI inference.

---

sgl is the node agent for the Singularity Layer decentralized compute grid. It runs a local llama.cpp inference server on your machine, pulls jobs from the grid orchestrator, decrypts end-to-end-sealed user prompts, runs inference, signs the result with your node key, and reports usage back for settlement. Operators stake $SGL to join the grid and earn a share of every compute payment.

• Grid: grid.x402compute.cc · Cloud: cloud.x402layer.cc
• Staking: staking.x402layer.cc — stake $SGL to register a node
• Staking program: Singularity-Layer/sgl-staking

Requirements

• macOS on Apple Silicon (M-series). Intel is unsupported; Linux/Windows land in Phase 2.
• 16 GB RAM minimum (more for larger models).
• llama.cpp — brew install llama.cpp.
• A GGUF model file (e.g. Llama 3.2 3B Instruct, Q4_K_M).
• A Solana wallet with staked $SGL (see staking.x402layer.cc).

Install

curl -sSf https://grid.x402compute.cc/install.sh | sh

Or build from source (see Build).

Quickstart

sgl detect                          # check your hardware (chip, RAM, GPU, Secure Enclave)
sgl login                           # browser device-auth → registers this node to your wallet
sgl start \
  --model-path ~/models/llama-3.2-3b-instruct-q4_k_m.gguf \
  --model-name llama-3.2-3b \
  --resource-percent 50             # dedicate ~50% of the machine to the grid

sgl login opens a browser device-authorization flow and binds the node to your staked wallet. sgl init --wallet <ADDRESS> is the non-interactive alternative.

Commands

Command	What it does
sgl detect	Show hardware capabilities (chip, cores, RAM, GPU, Secure Enclave, SIP).
sgl login	Browser device-auth login + register this node (recommended).
sgl init --wallet <ADDR>	Non-interactive: generate keys + register under a wallet.
sgl start --model-path <gguf> --model-name <name>	Start serving: heartbeat + process jobs.
sgl attest	Sign the orchestrator's challenge + send the hardware report.
sgl status	Node status, reputation, jobs completed/failed.
sgl service install | status | uninstall	Run as a background OS service across reboots.

sgl start flags include --resource-percent (a preset that scales threads / GPU layers / concurrent jobs), or fine-grained --threads, --gpu-layers, --context-size, --max-jobs, --batch-size, --inference-port, --heartbeat-interval. The global --orchestrator-url defaults to https://grid.x402compute.cc.

Run as a service (production)

sgl service install --model-path <gguf> --model-name <name> --resource-percent 50
sgl service status
sgl service uninstall

macOS uses launchd (wrapped in caffeinate -i to block idle sleep); Linux uses a systemd --user unit. Restarting reloads the model (~30–60 s) and briefly drops the node from the grid.

How confidentiality works

Prompts can be end-to-end encrypted between the caller and your node — the orchestrator relays ciphertext it cannot read:

1. Each node has an ed25519 identity key (crypto.rs). An X25519 key is deterministically derived from it (encryption.rs) and published on every heartbeat.
2. A caller seals the prompt to the node's X25519 key using X25519 + XChaCha20-Poly1305 (ephemeral sender key per message), and includes a response key.
3. The node decrypts in memory, runs inference locally, and seals the result back to the caller's response key. Only token-usage counts are sent in cleartext, so the grid can bill without seeing prompt or response content.
4. The node signs every result with its ed25519 key. A forged or altered result is detectable by the orchestrator and is grounds for slashing.

Integrity & anti-tamper

• Result signatures — every output is signed by the node key; the orchestrator verifies them and slashes nodes that return forged results.
• Binary allowlist — the orchestrator only accepts results from sgl builds whose sha256 is on a published allowlist, so a forked/tampered binary can't quietly serve.
• Self-reported hardware attestation (tee.rs) — the node reports SIP status, a hash of its own binary, and machine fingerprints on attest. This is a defense-in-depth deterrent, not a hardware root of trust: the report is produced by the node itself, so it raises the bar against casual tampering but does not cryptographically prove the execution environment the way a hardware-signed TEE quote (e.g. SGX/SEV/Apple DCAppAttest) would. Hardware-backed attestation is on the roadmap.
• Runtime hardening (runtime_hardening.rs) — anti-debugger (PT_DENY_ATTACH / prctl) to discourage live process tampering.
• Keys at rest — the node keypair is stored 0600 in a 0700 directory; the node warns if permissions are looser.

Confidentiality scope. Your prompts are protected in transit and are not persisted by the grid, but inference runs in your node's normal process memory. Treat this as strong transport + operational confidentiality, not hardware-enclave-isolated execution.

Build from source

Requires Rust (2021 edition) and llama.cpp.

git clone https://github.com/Singularity-Layer/sgl-network-node.git
cd sgl-network-node
cargo build --release
sudo cp target/release/sgl /usr/local/bin/
sgl detect

Security

Found a vulnerability? Please report it privately — open a GitHub security advisory on this repo rather than a public issue. The source is public for independent review; issues and PRs are welcome.

License

MIT © 2026 Singularity Layer