Managed text inference on one approved Mac per rental, allocated exclusively to one customer for 30 days. Choose a supported model and current machine, or supply your own Mac under reviewed terms.
For an AI agent: read /compute/skill.md first.
The full Markdown operating guide contains the supported commands, safe key handling, state checks and API reference. Paste this request into the assistant you use on your Mac:
Help me onboard or troubleshoot my Mac using https://agents.proxies.sx/compute/skill.md First inspect any existing installation and its current farmer dashboard. Check the prerequisites and diagnose the actual state before reinstalling. Keep keys and device tokens out of this chat; I enter secrets privately. Preserve the device identity and do not interrupt an existing rental. Explain what is verified, what is pending, and the next useful action.
An existing healthy service needs no reinstall. Setup installs a background service and model runtime; restarting, reinstalling or replacing keys on a rented Mac requires coordinated maintenance.
Start with a supported model, review compatible machines and confirm the selected node's current 30-day quote. Existing rentals are in Deployments.
| Tier | Memory minimum | Live default / 30 days |
|---|---|---|
| Starter | 24 GB | Check live quote |
| Pro | 48 GB | Check live quote |
| Max | 96 GB | Check live quote |
Admin controls defaults and node overrides. The individual checkout quote determines the purchase price. No per-token billing; request and model limits apply.
Check the marketplace for current availability before adding funds.
Use Spawn API key for the Compute supplier preset. Use Farmer Compute to inspect your Mac, review admission, list eligible capacity and see pricing/earnings.
Your Mac needs a current signed and countersigned Partner Agreement, a passing benchmark and explicit admin approval. Registration stages an unlisted device; installing does not automatically approve or rent it.
Successful paid rentals accrue the admin-approved farmer share. You can propose a price change; only admin can change effective pricing or the platform percentage. Complimentary allocations do not generate paid-rental share. You cover electricity, connectivity and hardware costs.
Requires native Apple Silicon, macOS 14+, at least 24GB RAM, native arm64 Node.js 18+ and Python 3.10+. Keep at least 26GB free for the smallest model; larger models need more space. See the full prerequisite checks.
node "$HOME/.proxies-compute/compute-agent.js" --status
Check the device ID, service and model against your farmer dashboard. Read the relevant recent log locally at ~/Library/Logs/proxies-compute.log. Do not run a second foreground agent or reinstall a working service. See credential recovery if authentication is the problem.
Sign in to the farmer portal, complete the Partner Agreement, then open Spawn API key. Select Compute supplier, enter a key name and choose Create API key. Save the one-time displayed secret privately before hiding it or leaving the page. This key grants account:read and compute:supply. In a separate working directory, inspect the downloaded script and run the installer in an interactive terminal:
node --version && python3 --version curl -fSLo compute-agent.js https://agents.proxies.sx/compute/compute-agent.js node compute-agent.js --install
Enter the complete key at the non-echoing prompt. Do not put it in chat, a command or a screenshot. The installer reuses a saved key, and an API_KEY environment value takes precedence; use the recovery guide if the selected key is wrong.
Agent 0.3.3 validates supplier access before changing the installation. It installs a private pinned Python/MLX environment, saves credentials with private permissions, copies itself to ~/.proxies-compute/compute-agent.js and loads the sx.proxies.compute launchd service. No global pip installation is required.
node "$HOME/.proxies-compute/compute-agent.js" --status node "$HOME/.proxies-compute/compute-agent.js" --doctor
Keep the Mac powered, awake and connected while its model downloads and loads. launchd starts at login; it does not prevent lid-close sleep. Match the device in the farmer portal and follow each remaining admission step.
| State | Meaning and next step |
|---|---|
| Offline / downloading / loading / paused | Address power, connectivity, disk or runtime progress. A heartbeat alone does not mean the model is ready. |
| Benchmark required | Keep a free Mac serving its model. An eligible heartbeat schedules a fresh measurement; eligibility is not proof a job is queued. |
| Benchmark pending | A job is queued or running. Wait for the outcome; it is separate from approval. |
| Awaiting approval | Keep the measured model ready, complete any agreement correction and wait for admin admission. |
| Blocked | An explicit admin review is required. Automatic heartbeats or passing diagnostics cannot remove a block. |
| Approved | Review the current price/share and listing control. Only ready, free, eligible listed machines are rentable. |
| Rented | Keep serving the customer. Cancelled-grace and outage-paused rentals still hold the machine. |
Admission uses a valid passing benchmark no more than 24 hours old, with its measured model still serving and a current countersigned agreement. Unapproved eligible Macs automatically refresh expired measurements. Approved suppliers retain a seven-day maintenance interval; a larger served model can require another benchmark. Automatic retries wait at least an hour and exclude occupied or admin-blocked Macs.
A benchmark measures output and throughput; it is not hardware attestation. The farmer/admin dashboards distinguish measurement, pending work and admission. Approval and listing do not guarantee demand or earnings.
One Mac serves a whole model from unified memory. The agent normally selects the largest supported model fitting its RAM; an existing renter's requested model takes precedence. Read the current pinned catalog.
| Model ID | Memory minimum | Context cap | Approximate weights |
|---|---|---|---|
| qwen3.8-27b-4bit | 24 GB | 16,384 tokens | 16 GB |
| qwen3.8-27b-8bit | 48 GB | 32,768 tokens | 30 GB |
| qwen3.8-27b-bf16 | 96 GB | 32,768 tokens | 55 GB |
The downloader checks the selected model's approximate weight size plus 10GB free space on the model-cache volume. A 48GB or 96GB Mac can therefore need more disk than the installer's 26GB minimum.
Agent 0.3.3 is readable Node.js source with no npm dependencies. It uses a private environment containing mlx-lm 0.31.3, MLX 0.31.2 and Transformers 5.4.0. The installed service retains its device identity across restarts.
| Operation | Behavior |
|---|---|
| Registration | Reports chip, RAM, cores, macOS and the Mac hostname as its node name. The agent manages device authentication after supplier registration. |
| Serving | Loads the catalog's pinned model with MLX on 127.0.0.1:8734. Checks a startup generation before reporting ready; refuses to kill an unrelated process using that port. |
| Heartbeat | Every 60 seconds: model state, served models, version and free disk. Three minutes without a valid heartbeat is offline. |
| Jobs | One job at a time; streams output and waits for the active job before a model switch. Invalidated rental claims stop generation. |
| Power | Checks every minute. Pauses below 20% battery when unplugged or below 50% thermal CPU limit; resumes when the condition clears. |
| Shutdown | Stops its MLX child and attempts an offline farewell heartbeat. Reinstalling replaces the existing service. |
| Removal | --uninstall pauses and keeps identity, runtime and models. --remove deletes the node on Proxies.sx and every file the agent created. See Pause or remove your Mac. |
Compute nodes do not carry proxy traffic. Renters receive managed inference access, not SSH, remote desktop, a shell or arbitrary containers.
Use the current downloaded script's safe diagnostic. It reports the selected key source and result code without printing the key:
node compute-agent.js --doctor env -u API_KEY node compute-agent.js --doctor
The second check excludes a shell override and checks the saved configuration. HTTP 401 means the selected key is not accepted; 403 means supplier role/scope access is missing. A successful check proves supplier API access, not a completed installation or passing benchmark.
When a replacement is needed on a free Mac, use node compute-agent.js --configure-key and enter the new farmer Compute key privately. It validates before saving and preserves identity/models; it does not restart the service. Follow its printed restart instruction only when safe. Read the full recovery guide; never delete identity files to force registration.
Your Mac is yours. You can pause or leave at any time, without asking us and without a notice period. The only thing that blocks removal is a customer renting the machine at that moment.
| Pause: --uninstall | Leave for good: --remove | |
|---|---|---|
| Your node on Proxies.sx | Stays in your dashboard, shown offline. Not rentable while offline. | Deleted with your own key: no longer listed, rentable or counted. |
| launchd service | Stopped and unloaded. | Stopped and unloaded. |
| Saved API key | Deleted from the Mac. | Deleted from the Mac. |
| Device identity | Kept, so a re-install is the same node. | Deleted. A later install is a new node. |
| Private Python/MLX runtime (about 1 GB) | Kept. | Deleted. |
| Downloaded catalog model (16–55 GB) | Kept. | Deleted. Other models in your Hugging Face cache are never touched. |
| Log ~/Library/Logs/proxies-compute.log | Kept. | Deleted. |
| Earnings already recorded | Unchanged. | Unchanged: they stay on your account and are paid as usual. |
node "$HOME/.proxies-compute/compute-agent.js" --uninstall
Needs agent 0.3.3 or newer. The first line fetches the current one, so this works whatever version is installed:
curl -fSLo /tmp/compute-agent.js https://agents.proxies.sx/compute/compute-agent.js node /tmp/compute-agent.js --remove
What it does, in order: it deletes the node on Proxies.sx through DELETE /v1/peer/my-devices/{deviceId} using the key saved on the Mac; stops the launchd service; then deletes ~/.proxies-compute (key, identity, runtime, the installed agent), ~/Library/LaunchAgents/sx.proxies.compute.plist, the log, and the catalog models it downloaded. Afterwards nothing from Proxies.sx runs on the Mac and nothing on it connects to us.
If the Mac is rented the command stops with a message and deletes nothing. Wait for the rental to end, or open a farmer ticket and we move the customer. If Proxies.sx cannot be reached it still cleans the Mac and tells you to delete the node under Compute in the farmer portal.
One step is left for you: revoke the API key under API keys in the farmer portal. A key cannot revoke itself. Keep it active if another Mac still uses it.
Prefer to do it by hand? Run launchctl bootout gui/$(id -u)/sx.proxies.compute, delete the files named above, then delete the node in the farmer portal.
Renters use their existing Proxies.sx account and USD balance. Add funds directly in the rental review or Compute billing using saved cards, a new Stripe card, or supported CoinGate crypto options. CoinGate checkout returns to Compute billing. Explicitly confirm the selected rental after the shared balance updates. Adding funds does not reserve capacity. Compute has no x402 checkout.
Admin sets effective prices and platform/farmer percentages, including node overrides. Farmers can submit a proposal; it takes effect only after admin approval. Accepted purchases retain their recorded terms. Renewal is manual at the newly reviewed quote.
Completed rental purchases are non-refundable; failed allocations are reversed to account balance. Cancellation preserves access through the term. Recorded unavailability pauses remaining rental time; customer-requested model downloads use rental time.
Farmer accrual, settlement and completed payout are separate. Inspect the authenticated earnings view and its settlement notice. Complimentary rentals generate no paid-rental share. Listing does not promise revenue.
POST https://api.proxies.sx/v1/peer/compute/rentals/<rentalId>/chat/completions Authorization: Bearer psx_<key> Required scope: compute:infer
The API accepts text system/user/assistant messages, temperature, max_tokens and optional SSE streaming. X-API-Key is also supported. The OpenAI SDK's api_key/apiKey option sends the Bearer header. See the renter examples and scope/route reference.
No per-token billing. Limits include one running job per node, 1–4,096 output tokens (default 1,024), catalog context and a conservative input-byte budget. A 60 requests/minute endpoint throttle applies by client IP/handler, not independent capacity per key or rental. Additional per-key/account limits may apply. The API's default deadline is three minutes; the agent stops generation at 175 seconds or after 60 seconds without output.
Tool calling, response_format, image/audio input, arbitrary models, training, fine-tuning, customer containers and multi-Mac clusters are unsupported. For RAG, your application retrieves documents and sends the relevant text context.
No. Current demand and eligible stock determine rentals. Verify your own machine's agreement, runtime, admission, price and allocation state. A local test or an installation is not a paid customer outcome.
No. Suppliers own and administer their Macs, including their operating system and memory. Customer prompts pass through the platform and supplier runtime. The agent does not intentionally persist prompts, but runtime errors can contain sensitive data; inspect and redact diagnostic excerpts before sharing.
Completed prompts/results are normally scrubbed after one hour and job records expire after 24 hours, subject to cleanup timing. This is not zero retention. Never publish configuration, device-token files, authorization headers or customer prompts.
Yes, at any time and without asking us. --uninstall pauses; --remove takes the Mac off the network for good and deletes everything the agent created. The full list of what each one keeps and deletes is under Pause or remove your Mac.
Open a farmer dashboard ticket with device ID, agent version, OS/chip/RAM, observation time, safe doctor result, model/admission/pending state and a short redacted error. Say whether a customer rental is affected. Early access has no availability SLA.
AI-assistant operating guide · Machine-readable facts · Renter API guide · Compute provider fit checklist · Agent source
Customer models · Customer deployments · Farmer Compute · Farmer API keys · Current rentable machines · Ecosystem