Gpusmarket

gpusmarket

Rent real GPUs from the terminal. Built for AI agents.

[npm version] [MCP — 32 tools] [Agent-native] [smithery badge]

One command to sign up, one to rent, one to run inference. Per-second billing, no subscriptions, $0.50 minimum per rental. The CLI your coding agent can drive end-to-end: GPUsMarket.

npm install for GPUs. One command in, one command out.

Renting a GPU usually means a signup form, a quota request, a support ticket, and a $50 minimum — before you've run a single token. GPUsMarket collapses that to three commands: everything from consumer RTX 4090s to datacenter H100s and MI300Xs, rentable by the second. And it's built so your agent can sign up and rent by itself — no human, no browser, no dashboard.

Works with: Claude Code · Cursor · Cline · Windsurf · Aider · Codex · any MCP client

[gpusmarket demo]

Install to real inference in under a minute — watch the full-res demo.

npm install -g gpusmarketgpusmarket init

Quick start

# Create an account (returns your API key, auto-saved)gpusmarket signup --email [email protected] --tenant my-agent
# Find a GPUgpusmarket search --gpu "RTX 4090"
# Rent the cheapest match — returns an OpenAI/Ollama-compatible endpoint + keygpusmarket rent --gpu "RTX 4090"
# Run inference straight through your rentalgpusmarket chat "What is 2+2?" --model gemma3:4b
# Stop when done — you pay for exactly the seconds you usedgpusmarket stop RENTAL-ID

Built for agents

Built so an agent can go from zero to running inference with no human in the loop — sign up, rent, infer, and stop, all from the terminal or over MCP. This CLI is designed to be driven by AI coding agents (Claude Code, Cursor, Cline, Windsurf, Codex, Copilot), not just humans.

• gpusmarket init — in-terminal quickstart: detects your agent's config file, prints a ready-to-paste skill block, MCP setup, and next steps.
• gpusmarket init --agent-schema — the full command contract as JSON: every command, every flag, plus the rules agents must follow. Load it into context and your agent never hallucinates a flag.
• --json everywhere — every command has machine-readable output.
• Zero dependencies — a single self-contained Node.js CLI. Nothing to audit but us.

Watch an agent drive it end-to-end (click for the full-res video):

[agent drives gpusmarket via MCP]

Claude Code searches, rents, runs inference, and stops the GPU — no human in the loop.

MCP server

gpusmarket mcp-serve exposes 32 tools over stdio — search, rent, status, stop, chat, host, pools, the whole surface. Add it to Claude Code:

claude mcp add gpusmarket -- gpusmarket mcp-serve

…or drop it into any MCP client's .mcp.json. The MCP server runs outside your project directory, so pass your API key via the GPUSMARKET_API_KEY environment variable:

{  "mcpServers": {    "gpusmarket": {      "type": "stdio",      "command": "gpusmarket",      "args": ["mcp-serve"],      "env": { "GPUSMARKET_API_KEY": "gpu_your_key_here" }    }  }}

No key yet? Run gpusmarket signup to provision one, or grab it from your dashboard.

Core capabilities

Each demo links to the full-resolution MP4.

# Browsegpusmarket search --gpu "RTX 4090"      # filter by GPU, price, modelgpusmarket cheapest --model gemma3:4b   # the single cheapest node that serves a modelgpusmarket models                       # every model available across the marketplace
# Managegpusmarket rentals                      # your active rentalsgpusmarket status RENTAL-ID             # live status + running costgpusmarket stop RENTAL-ID               # stop, pay for the seconds you used
# Pool — one endpoint, several rentals behind itgpusmarket pool create my-poolgpusmarket pool add my-pool RENTAL-IDgpusmarket pool list

Honest billing

• Per-second billing. A 90-second experiment costs 90 seconds of GPU time.
• A hold, not a charge. Renting places a pre-auth hold (default $10) — it's released in full when you stop. You are never charged the hold.
• $0.50 minimum per rental. You pay per second for the compute you consume, subject to a $0.50 minimum charge per rental (disclosed when you rent). When you stop, your card is charged the exact usage cost, or the $0.50 minimum — whichever is greater.
• Receipts for everything, with a full "where your money goes" breakdown.

Earn with your GPU

Your GPU earns nothing while it sits idle. List it in one command — the CLI auto-detects your hardware, prints your host key, and the listing goes live in minutes. You keep 90% of every rental, paid out to your Stripe account automatically. No port forwarding, no static IP, no open inbound ports.

Three ways to host (click any demo for the full-res video):

gpusmarket host setup          # auto-detects your GPU, creates the listingnpx gpusmarket-host --key KEY  # connect your machine — no port forwarding needed

Commands

Run gpusmarket --help for the full reference, flags included.

Features

• Search — filter the marketplace by GPU, price, or the exact model you need to run.
• Rent — one command returns an OpenAI/Ollama-compatible endpoint + key, ready for inference.
• Chat — stream inference straight through your rental, or pipe from stdin.
• Pools — put several rentals behind one endpoint and load-balance across them.
• Host — list your own GPU (Mac, NVIDIA, or Docker) and keep 90% of every rental.
• Per-second billing — receipts for everything, $0.50 minimum per rental, no subscriptions.
• MCP server — 32 tools over stdio for Claude Code, Cursor, Cline, Windsurf, any MCP client.
• Agent-native — --json everywhere plus init --agent-schema so agents never guess a flag.

Why this exists

Renting a GPU should be as fast as npm install. Not a signup form, a quota request, a support ticket, and a $50 minimum. GPUsMarket is a marketplace where anyone can list an idle GPU and anyone — or any agent — can rent it by the second. This CLI is the agent-native front door to that marketplace. It's early and moving fast: if something's rough or missing, open an issue — we read every one.

Documentation

Full docs at gpusmarket.com. Run gpusmarket init for the interactive quickstart.

License

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