Hatch turns idle Mac Minis and Mac Studios into a global GPU network. Lock your screen and earn HC credits. Pay cents to run 70B LLMs on 192 GB of unified memory.
curl -fsSL https://raw.githubusercontent.com/wkang0223/hatch/master/scripts/install-agent-macos.sh | bash
Unified memory means the GPU and CPU share one pool. No copies. Full bandwidth. A $1,400 Mac Mini M4 Pro outperforms a $30,000 pair of A100s for 70B LLM inference.
| Machine | Unified RAM | GPU Cores | Best For | Price |
|---|---|---|---|---|
| Mac Mini M4 | 16 – 24 GB | 10 | 7B – 13B inference, agents | $599 |
| Mac Mini M4 Pro Best value | 24 – 64 GB | 20 | 30B – 70B LLM inference | $1,399 |
| Mac Studio M4 Max | 64 – 128 GB | 40 | 70B fine-tuning, multi-modal | $1,999 |
| Mac Studio M4 Ultra | 128 – 192 GB | 80 | 405B shards, full pipeline | $3,999 |
Three commands and your Mac is earning. Jobs run in an isolated macOS sandbox — you never lose control.
One-line installer. Creates the hatch_worker isolation user, installs MLX & PyTorch MPS runtimes, and registers a launchd daemon that starts on boot.
The agent polls GPU utilization via IOKit every 30 s. Once GPU stays below 5% for 10 minutes with your screen locked, your Mac announces itself as available.
The coordinator matches jobs to your Mac by RAM requirement and runtime. Jobs execute inside a sandbox-exec profile — file and network access fully restricted.
Earnings settle instantly in HC (Hatch Credits). Cash out to USDC on Arbitrum or spend credits to run your own compute jobs.
No existing platform — io.net, Vast.ai, Salad, Akash — supports Apple Silicon. Hatch fills that gap.
Uses macOS CGSession API. Your GPU only leases when your screen is locked and GPU is truly idle.
Reads IOAccelStatistics2 every 30 s. Drops off the network the instant your GPU spikes above threshold.
Jobs run as a restricted hatch_worker OS user with sandbox-exec profiles — no filesystem or outbound network beyond the job directory.
Full Metal GPU access. MLX and PyTorch MPS backends run at native speed — zero virtualization overhead.
Phase 1: signed Ed25519 escrow receipts in PostgreSQL. Phase 3: trustless settlement on Arbitrum L2.
Agent, coordinator, and ledger are all written in Rust. Single binary per component — no runtime dependencies.
Use the live network, self-host on Railway, or fork and build your own.
Install the Hatch agent on your M-series Mac. One command — detects your chip, installs runtimes, sets up launchd, and starts earning.
curl -fsSL https://raw.githubusercontent.com/wkang0223/hatch/master/scripts/install-agent-macos.sh | bash
Self-host the coordinator + ledger on Railway in minutes. Click deploy, add a Postgres database, set three env vars, and you have your own private Hatch network.
Full Rust monorepo — agent, coordinator, ledger, CLI, Next.js dashboard. Apache 2.0 licensed. Build your own GPU marketplace on top.
cargo build -p hatch-agentcargo build -p neuralmesh-coordinatorcargo build -p hatch