First action - register your Mac now (one command, below). Registered nodes are benchmarked and first in line when serving and payouts switch on.
1. Create a free account + API key at farmer.proxies.sx (same key as the Peer Network).
2. Register your Mac:
curl -O https://agents.proxies.sx/compute/compute-agent.js API_KEY=psx_your_key node compute-agent.js
The agent detects your chip and unified memory via sysctl, registers the node, and heartbeats. Your Mac appears live at farmer.proxies.sx/compute within a minute.
3. Serve a model (v0.2 — jobs flow end-to-end today):
pip install mlx-lm API_KEY=psx_your_key SERVE_MODEL=qwen3.8-27b-4bit node compute-agent.js
The agent runs the pinned model under MLX (Apple's inference runtime), pulls inference jobs, and posts results. Pick a model your RAM covers from the live catalog: GET /v1/peer/compute/catalog. First start downloads the model from Hugging Face (15GB+ for 27B-class).
Requires macOS on Apple Silicon (M1 or newer) and Node 18+; serving additionally needs Python 3 + mlx-lm. Desktop Macs (mini / Studio) are the ideal nodes — always powered, actively cooled. MacBook Pro works; fanless MacBook Air throttles under sustained load and is not recommended. Serving works today and earns from launch day — nothing accrues in early access.
A marketplace where Apple Silicon Macs are rented as DEDICATED, managed inference endpoints — one customer per machine, unlimited tokens, OpenAI-compatible API. Your Mac's unified memory is the asset: a 64–128GB Mac loads models that would otherwise need a datacenter GPU. The customer side lives at compute.proxies.sx.
How you earn: register your Mac (below), get verified, then list it for rent from farmer.proxies.sx/compute. Listed machines appear anonymized on the public marketplace, priced by memory tier ($249–$2,499/mo). While rented, you earn 45% of the monthly price — e.g. an M4 Max 128GB rents at $799/mo → ~$360/mo to you, vs ~$10/mo of electricity. Payouts ride the same rails that already pay Peer Network farmers.
What the customer can NEVER do: touch your machine. No SSH, no files, no shell — only the managed inference API reaches your Mac, and it serves only the pinned catalog model you can see. Compute nodes also never carry proxy traffic — excluded from proxy routing by design.
Each node serves whole models from its own unified memory. What fits (4-bit quantized unless noted):
| Mac | Unified memory | Serves (examples) |
|---|---|---|
| M4 / M4 Pro mini | 16–24 GB | 7B–14B class, 27B at reduced context |
| M4 Pro / Max | 36–64 GB | 27B–32B full context, 70B (4-bit) |
| M4 Max | 128 GB | 70B comfortably, 27B at bf16, large-context tiers |
| M3 Ultra Studio | 256–512 GB | DeepSeek-671B-class MoE (4-bit), trillion-param MoE tiers |
The model catalog is curated: pinned models, pinned quantizations, verified by hash — so customers get exactly the model they paid for and nodes are verified by reproducible benchmark probes.
Have several Macs in one place? Wire them with Thunderbolt and register them as one node with summed memory:
CLUSTER_SIZE=4 RAM_GB=512 API_KEY=psx_your_key node compute-agent.js
Local Thunderbolt fusion (80 Gb/s, microsecond latency) is the only place model-sharding works — that's physics, not policy. Model splitting over home internet adds 15–25ms per pipeline hop per token and collapses throughput, so the marketplace never shards a model across the WAN. Scale comes from many whole nodes serving many requests, with per-session routing — the same way every production inference fleet works.
45% of the rental price while your machine is rented — the tier ladder is $249 / $449 / $799 / $2,499 per 30 days by memory class (24–36GB / 48–64GB / 128GB / 256–512GB), so a rented machine pays you roughly $112 / $202 / $360 / $1,125 a month. An unrented listed machine earns nothing and costs you nothing — no demand-risk transferred to you. Rental demand is early; listing is how you're first in line for it.
Customers rent from what's visible. Verified, listed machines appear on the public marketplace immediately, and capacity is assigned listing-first. Registration costs one command and ~0 resources (a heartbeat every 60s).
Today: this open agent script (read it — it's ~150 lines). At launch: a signed, notarized macOS app running curated open models under MLX, with pause-on-battery, pause-on-user-activity, and thermal backoff. You approve each release; nothing else executes.
Registration sends: chip name, RAM, core count, hardware model, macOS version, and your API key. No file access, no browsing data, no proxy traffic — ever.
compute-agent.js (the agent, readable source) · your compute dashboard · Peer Network (earn with bandwidth today) · agents.proxies.sx