gpusmarket-host

List your GPU on GPUsMarket in one command.

Auto-detects your GPU, installs the inference backend, pulls the best model for your hardware, and connects to the marketplace. You start earning.

Quick Start

npx gpusmarket-host --key gph_your_key

Get your host key at gpusmarket.com/host/setup.

What It Does

1. Detects your GPU (NVIDIA, Apple Silicon, AMD)
2. Installs Ollama or llama.cpp (if not present)
3. Pulls the best model for your VRAM
4. Starts the inference server
5. Connects a secure tunnel to GPUsMarket
6. Your GPU is live and earning

Options

--key <key>         Host key from gpusmarket.com/host/setup (required)
--workload <type>   llm (default), audio, image, or video
--backend <type>    ollama (default) or llama-cpp   (llm only)
--model <name>      Specific model (auto-recommended if omitted)
--target <url>      Tunnel an already-running server (skip install/pull/serve)
--port <port>       Server port (default per workload)
--preview           After the tunnel is live, run one sample on your own GPU
                    and open/play the result (instant proof it works)
--install           Stay online across reboots (launchd/systemd service)
--uninstall         Remove the persistent service
--dry-run           Show plan without executing
--no-install        Skip backend installation
--no-pull           Skip model download

Examples

# Default — Ollama, auto-detect everything
npx gpusmarket-host --key gph_abc123

# Use llama.cpp for better raw performance
npx gpusmarket-host --key gph_abc123 --backend llama-cpp

# Specific model
npx gpusmarket-host --key gph_abc123 --model qwen3:32b

# llama.cpp with specific GGUF
npx gpusmarket-host --key gph_abc123 --backend llama-cpp --model Qwen/Qwen3-32B-GGUF:qwen3-32b-q4_k_m.gguf

# Tunnel an Ollama you already run yourself (separate container, remote box, etc.)
npx gpusmarket-host --key gph_abc123 --target http://localhost:11434

Host more than LLMs (--workload)

Same one command, same tunnel — only the model server behind it changes. The marketplace, metering and payouts are workload-blind, so any of these just work:

# Audio — Kokoro-82M TTS + whisper-small STT (OpenAI /v1/audio/* contract).
# Tiny + fast: runs on ANY Apple-silicon Mac (Metal/MPS) or NVIDIA GPU.
npx gpusmarket-host --key gph_abc123 --workload audio --preview

# Image — SDXL-Turbo (OpenAI /v1/images/generations). Mac (MPS) or NVIDIA.
npx gpusmarket-host --key gph_abc123 --workload image --preview

# Video — Wan2.1-T2V async job contract. NVIDIA only.
npx gpusmarket-host --key gph_abc123 --workload video --preview

The first run of a non-LLM workload builds a small Python venv (uses uv if present) and lazily downloads the model weights on first use; later runs reuse it and start fast. Add --preview to have your own GPU produce a sample image/clip/voice line the moment the tunnel is live — the file lands in ~/.gpusmarket/preview/ and opens automatically.

Bring your own server (--target)

If you already run Ollama (or any OpenAI/Ollama-compatible server) — for example in its own Docker container or on another machine — point gpusmarket-host at it with --target and it skips install/pull/serve entirely, connecting the tunnel straight to your endpoint:

npx gpusmarket-host --key gph_abc123 --target http://my-ollama:11434

This is the recommended setup for Docker/Compose hosters who keep inference in a dedicated container.

Stay online across reboots (--install)

By default gpusmarket-host runs in the foreground — close the terminal or reboot and your GPU goes offline. To keep it listed and earning around the clock, install it as a background service:

npx gpusmarket-host --key gph_abc123 --install

This registers a boot-persistent service that re-runs your exact command (same key(s), backend, model, target) and restarts it if it ever exits:

• macOS — a per-user LaunchAgent (~/Library/LaunchAgents), RunAtLoad + KeepAlive
• Linux — a systemd --user unit (~/.config/systemd/user), Restart=always + linger (starts at boot without login)

All your normal flags work with --install, including parallel --keys and --target:

npx gpusmarket-host --key gph_abc123 --backend llama-cpp --install

Check what it would do first with --dry-run, and remove it any time:

npx gpusmarket-host --key gph_abc123 --install --dry-run   # preview, no changes
npx gpusmarket-host --uninstall                            # stop & remove

Logs live at ~/Library/Logs/gpusmarket-host/agent.log (macOS) or journalctl --user -u gpusmarket-host.service -f (Linux).

Requirements

• Node.js 18+
• macOS or Linux
• A GPU (NVIDIA, Apple Silicon, or AMD)

License

Proprietary - Copyright (c) 2026 Tyga.Cloud Ltd. All rights reserved. See LICENSE file.