free-gpu 0.1.5

Plan local and free-tier GPU workflows around llmfit and curated provider data.

free-gpu

Plan local-first experiments and get more from free GPU providers.

See what fits on your hardware first, then use free or near-free GPU providers for the runs that should not stay local.

Website | PyPI | Dataset | Hosted MCP

---

free-gpu sits on top of llmfit and maps a workload to a practical path: stay local when your machine can carry it, then move to free tiers, credits, or grant-style compute only when the workload really needs it.

Table of Contents

• What it does
• Add to your MCP client
• Quick start
• How it works
• Provider data
• Project links

What it does

• Checks whether a workload is realistic on your local hardware through llmfit.
• Ranks provider lanes across free, under-25, and grant budgets.
• Plans common workloads such as inference, finetune-lora, batch-eval, and agent-loop.
• Exposes the same planner through a CLI, a TUI, a local MCP server, and a hosted MCP endpoint.

Add to your MCP client

Codex

Hosted:

codex mcp add freeGpu --url https://free-gpu.vercel.app/mcp

Local:

codex mcp add free-gpu-local -- free-gpu-mcp

Claude Code

Hosted:

claude mcp add --transport http free-gpu https://free-gpu.vercel.app/mcp

Local:

claude mcp add --transport stdio free-gpu -- free-gpu-mcp

Cursor

Hosted:

{
  "mcpServers": {
    "free-gpu": {
      "url": "https://free-gpu.vercel.app/mcp"
    }
  }
}

Local:

{
  "mcpServers": {
    "free-gpu": {
      "command": "free-gpu-mcp"
    }
  }
}

VS Code

Hosted:

{
  "servers": {
    "freeGpu": {
      "type": "http",
      "url": "https://free-gpu.vercel.app/mcp"
    }
  }
}

Local:

{
  "servers": {
    "freeGpu": {
      "type": "stdio",
      "command": "free-gpu-mcp"
    }
  }
}

Quick start

Install

pip install free-gpu

Open the TUI

free-gpu ui

Ask the planner from the CLI

free-gpu providers --workload inference --budget free
free-gpu plan --workload finetune-lora --model llama-3.1-8b --budget under-25 --task-hours 6 --min-vram-gb 16
free-gpu plan --workload scratch-train --budget grant --task-hours 24 --min-vram-gb 40

The LoRA fine-tuning example above is the core workflow: describe the model, define the workload, add a budget lane, and let the planner narrow down the realistic path.

How it works

1. Start with the workload shape: model, hours, VRAM target, and budget lane.
2. Check whether the run is realistic on your own machine through llmfit.
3. Rank providers only after the local fit is known.
4. Return a practical next step instead of making you browse pricing pages and free-tier docs manually.

Provider data

The provider ledger lives in free_gpu/gpu_compute_database.csv.

Project links

• Repository: https://github.com/francescoopiccolo/free-gpu
• Website: https://francescoopiccolo.github.io/free-gpu/
• PyPI: https://pypi.org/project/free-gpu/
• Dataset: free_gpu/gpu_compute_database.csv
• Hosted MCP: https://free-gpu.vercel.app/mcp