Website · 1-minute video · Agents guide · Wiki · Changelog

Your agent rents the GPU it needs, on its own. You pay only while it works: we watch every machine and shut down the idle ones.

Ask Claude Code, Codex, Gemini CLI or Cursor to "compare Qwen 7B and Llama 8B on these 50 prompts". The agent finds a GPU under your budget, serves the model with vLLM behind an OpenAI-compatible API, and runs the experiment. A watchdog on the machine shuts it down after idle_timeout (15 minutes by default) with nothing using it, even if the agent forgot, the chat closed or your laptop is asleep.

The agent installs it from one pasted line. A skill and local MCP tools (gpu_offers, gpu_setup, gpu_run, gpu_status, gpu_destroy, ...) cover the whole lifecycle, and the skill tells the agent to show you the price and ask before renting. Budgets, a maximum lifetime and an account-wide sweep catch anything left behind. Apps in any language get the endpoint from gpu --json setup.

Vast.ai is built in. Any other GPU provider takes one adapter file, and your agent can write it: see Use any GPU provider.

The quickest way works in every agent. Paste this, naming the agent you use:

Set up GPU for Agents for me. I'm using Claude Code. Fetch https://gpu4agents.jambu.ai/install.md with curl and follow its steps.

The agent installs the gpu command (with uv, which brings its own Python), the gpu4agents skill and the gpu MCP tools. It rents nothing.

You can also install the skill the way your agent installs other skills. The skill installs the gpu command the first time it runs.

/plugin marketplace add jambuai/gpu4agents
/plugin install gpu4agents@gpu4agents

On Claude Code 2.1.275 or newer, one line does both: /plugin install gpu4agents --marketplace jambuai/gpu4agents.

npx skills add jambuai/gpu4agents --global --agent codex

npx skills add jambuai/gpu4agents --global --agent '*'

This installs the skill for every agent the Skills CLI supports. Leave off --global to install it only in the current project.

Download gpu4agents-skill.zip and upload it under Settings > Capabilities > Skills.

curl -fsSL https://gpu4agents.jambu.ai/install.sh | sh              # uv, Python and gpu, then every agent it finds
uv tool install gpu4agents && gpu agent install -t auto --global    # the same, by hand

Run this in your own terminal, so the agent never sees the key:

gpu login                     # Vast.ai: paste a key from https://cloud.vast.ai/manage-keys/
gpu login --provider runpod   # a provider you added an adapter for
gpu login --check             # which account gpu finds, and its balance

gpu login checks the key with the provider, then saves it in ~/.gpu/credentials (mode 600). Keys in the environment or a project .env work too. They never go in jambu.yaml.

Restart the agent and ask for an experiment. See Letting a coding agent rent GPUs.

gpu talks to providers through a small contract: find offers, rent, report status, stop, destroy, and two calls the watchdog makes from the machine to shut itself down. Vast.ai ships with it. For anything else, such as RunPod, Lambda or your company's cluster, paste this into your agent:

Add <provider> to GPU for Agents. Fetch https://gpu4agents.jambu.ai/providers.md with curl and follow its steps.

The agent runs gpu providers new <name>, fills in the adapter from the provider's API docs, checks it with gpu validate and gpu offers, and asks you before a test rental. The adapter is one file in your project, selected in jambu.yaml:

gpu_runtime:
  provider:
    name: runpod
    plugin: ./gpu_providers/runpod.py:RunpodProvider

To reuse an adapter across projects, publish it as a package with a gpu4agents.providers entry point. The watchdog, budgets, CLI and MCP tools work the same with every provider. See PROVIDERS.md and Adding a provider.

gpu init --model Qwen/Qwen2.5-7B-Instruct
gpu validate
gpu offers                         # see prices, rent nothing
gpu setup                          # rent, boot vLLM, wait until the model answers
gpu run python experiment.py       # OPENAI_BASE_URL / OPENAI_API_KEY are set for the job
gpu destroy                        # or walk away: the idle timeout does it

From an application in any language, gpu --json setup prints the endpoint as JSON. See Integrating into an application, with a ready TypeScript client.

# jambu.yaml (only the gpu_runtime key is read, so other Jambu tools can share the file)
gpu_runtime:
  provider: {name: vast}
  model_profile: qwythos-9b
  lifecycle:
    idle_timeout: 15m          # the watchdog on the machine stops it
    max_lifetime: 6h           # a hard ceiling, busy or not
  budget:
    max_hourly_cost_usd: 0.60  # offers above this are never rented
    max_session_cost_usd: 5.00
  health:
    pull_stall_timeout: 10m    # a host whose image pull stalls is replaced

A full example is jambu.example.yaml. gpu ls shows every machine billing in the account, and gpu sweep --install-timer guards every project on the machine and cleans up orphans. See Keeping the account clean.

• Letting a coding agent rent GPUs (Claude Code, Codex, Gemini CLI, Cursor, Claude Desktop)
• Integrating into an application (any language)
• Commands, workloads, and agents
• Every config field
• Choosing a model
• How shutdown works
• Adding a provider, and the steps for an agent

The package was called jambu-gpu until 0.4.1. The code, the gpu command, jambu.yaml and the machines are the same:

pipx uninstall jambu-gpu && pipx install gpu4agents

Sweep timers installed by jambu-gpu keep working; gpu sweep --install-timer rewrites them for the new module.

Issues and pull requests are welcome. Read CONTRIBUTING.md and the Code of Conduct. Report security issues privately, as described in SECURITY.md.

Created by João Gabriel Lima at Jambu.ai.

MIT.