gpu-cliRun ML training, LLM inference, and ComfyUI workflows on remote NVIDIA GPUs (A100, H100, RTX 4090). Cloud GPU compute with smart file sync — prefix any comma...
Install via ClawdBot CLI:
clawdbot install angusbezzina/gpu-cliGPU CLI runs local commands on remote NVIDIA GPUs by prefixing with gpu. It provisions a pod, syncs your code, streams logs, and syncs outputs back: uv run python train.py becomes gpu run uv run python train.py.
gpu doctor --json # Check if setup is healthy (daemon, auth, provider keys)
gpu status --json # See running pods and costs
gpu inventory --json # See available GPUs and pricing| Command | Purpose |
|---|---|
| gpu login | Browser-based authentication |
| gpu logout [-y] | Remove session |
| gpu init [--gpu-type T] [--force] | Initialize project config |
| gpu upgrade | Open subscription upgrade page |
| Command | Purpose |
|---|---|
| gpu run | Execute on remote GPU (main command) |
| gpu run -d | Run detached (background) |
| gpu run -a | Reattach to running job |
| gpu run --cancel | Cancel a running job |
| gpu status [--json] | Show project status, pods, costs |
| gpu logs [-j JOB] [-f] [--tail N] [--json] | View job output |
| gpu attach | Reattach to job output stream |
| gpu stop [POD_ID] [-y] | Stop active pod |
Key gpu run flags: --gpu-type, --gpu-count <1-8>, --min-vram, --rebuild, -o/--output, --no-output, --sync, -p/--publish , -e , -i/--interactive.
| Command | Purpose |
|---|---|
| gpu inventory [--available] [--min-vram N] [--max-price P] [--json] | List GPUs with pricing |
| Command | Purpose |
|---|---|
| gpu volume list [--detailed] [--json] | List network volumes |
| gpu volume create [--name N] [--size GB] [--datacenter DC] | Create volume |
| gpu volume delete | Delete volume |
| gpu volume extend | Increase size |
| gpu volume set-global | Set default volume |
| gpu volume status [--volume V] [--json] | Volume usage |
| gpu volume migrate | Migrate to datacenter |
| gpu volume sync | Sync between volumes |
| Command | Purpose |
|---|---|
| gpu vault list [--json] | List encrypted output files |
| gpu vault export | Export decrypted file |
| gpu vault stats [--json] | Storage usage stats |
| Command | Purpose |
|---|---|
| gpu config show [--json] | Show merged config |
| gpu config validate | Validate against schema |
| gpu config schema | Print JSON schema |
| gpu config set | Set global config option |
| gpu config get | Get global config value |
| Command | Purpose |
|---|---|
| gpu auth login [--profile P] | Authenticate with cloud provider |
| gpu auth logout | Remove credentials |
| gpu auth status | Show auth status |
| gpu auth add | Add hub credentials (hf, civitai) |
| gpu auth remove | Remove hub credentials |
| gpu auth hubs | List configured hubs |
| Command | Purpose |
|---|---|
| gpu org list | List organizations |
| gpu org create | Create organization |
| gpu org switch [SLUG] | Set active org context |
| gpu org invite | Invite member |
| gpu org service-account create --name N | Create service token |
| gpu org service-account list | List service accounts |
| gpu org service-account revoke | Revoke token |
| Command | Purpose |
|---|---|
| gpu llm run [--ollama\|--vllm] [--model M] [-y] | Launch LLM inference |
| gpu llm info [MODEL] [--url URL] [--json] | Show model info |
| Command | Purpose |
|---|---|
| gpu comfyui list [--json] | Browse available workflows |
| gpu comfyui info | Show workflow details |
| gpu comfyui validate | Pre-flight checks |
| gpu comfyui run | Run workflow on GPU |
| gpu comfyui generate " | Text-to-image generation |
| gpu comfyui stop [WORKFLOW] [--all] | Stop ComfyUI pod |
| Command | Purpose |
|---|---|
| gpu notebook [FILE] [--run] [--new NAME] | Run Marimo notebook on GPU |
Alias: gpu nb
| Command | Purpose |
|---|---|
| gpu serverless deploy [--template T] [--json] | Deploy endpoint |
| gpu serverless status [ENDPOINT] [--json] | Endpoint status |
| gpu serverless logs [ENDPOINT] | View request logs |
| gpu serverless list [--json] | List all endpoints |
| gpu serverless delete [ENDPOINT] | Delete endpoint |
| gpu serverless warm [--cpu\|--gpu] | Pre-warm endpoint |
| Command | Purpose |
|---|---|
| gpu template list [--json] | Browse official templates |
| gpu template clear-cache | Clear cached templates |
| Command | Purpose |
|---|---|
| gpu daemon status [--json] | Show daemon health |
| gpu daemon start | Start daemon |
| gpu daemon stop | Stop daemon |
| gpu daemon restart | Restart daemon |
| gpu daemon logs [-f] [-n N] | View daemon logs |
| Command | Purpose |
|---|---|
| gpu dashboard | Interactive TUI for pods and jobs |
| gpu doctor [--json] | Diagnostic checks |
| gpu agent-docs | Print agent reference to stdout |
| gpu update [--check] | Update CLI |
| gpu changelog [VERSION] | View release notes |
| gpu issue ["desc"] | Report issue |
| gpu desktop | Desktop app management |
| gpu support | Open community Discord |
gpu login then gpu initgpu run python train.py --epochs 10gpu run --gpu-type "RTX 4090" python train.pygpu run -d python long_training.py then gpu status --jsongpu status --jsongpu logs --jsongpu stop -ygpu llm run --ollama --model llama3 -ygpu comfyui run flux_schnellgpu doctor --jsongpu run is pod-reuse oriented: after a command completes, the next gpu run reuses the existing pod until you gpu stop or cooldown ends.
Most commands support --json for machine-readable output. Structured data goes to stdout; human-oriented status and progress messages go to stderr.
Commands with --json: status, logs, doctor, inventory, config show, daemon status, volume list, volume status, vault list, vault stats, comfyui list, comfyui info, comfyui validate, serverless deploy, serverless status, serverless list, template list, llm info.
| Code | Meaning | Recovery |
|---|---|---|
| 0 | Success | Proceed |
| 1 | General error | Read stderr |
| 2 | Usage error | Fix command syntax |
| 10 | Auth required | gpu auth login |
| 11 | Quota exceeded | gpu upgrade or wait |
| 12 | Not found | Check resource ID |
| 13 | Daemon unavailable | gpu daemon start, retry |
| 14 | Timeout | Retry |
| 15 | Cancelled | Re-run if needed |
| 130 | Interrupted | Re-run if needed |
gpu.toml, gpu.jsonc, or pyproject.toml [tool.gpu]~/.gpu-cli/config.toml (via gpu config set/get).gitignore controls upload; outputs patterns control downloadGPU_RUNPOD_API_KEY, GPU_SSH_PRIVATE_KEY, GPU_SSH_PUBLIC_KEYreferences/create.mdreferences/debug.mdreferences/config.mdreferences/optimize.mdreferences/volumes.mdGenerated Mar 1, 2026
Data scientists and ML engineers can train models like PyTorch or TensorFlow on high-end NVIDIA GPUs (e.g., A100, H100) without investing in local hardware. The tool syncs code and data to remote pods, runs training scripts via commands like 'gpu run uv run python train.py', and streams logs and outputs back, enabling scalable training for tasks such as image classification or natural language processing.
Developers can deploy large language models for inference using commands like 'gpu llm run' with options for Ollama or vLLM backends. This allows real-time text generation, chatbots, or API endpoints on powerful GPUs, with features to manage endpoints, monitor logs, and scale resources as needed, ideal for startups or enterprises building AI-powered applications.
Artists and designers can run ComfyUI workflows for stable diffusion and text-to-image generation via commands like 'gpu comfyui generate'. This enables high-quality image creation, video processing, or other GPU-intensive creative tasks on remote RTX 4090 or similar GPUs, with tools to manage workflows, validate setups, and sync outputs securely.
Researchers in academia can prototype AI models or run experiments on-demand without managing infrastructure. Commands like 'gpu notebook' allow running Marimo notebooks remotely, while features for volumes and encrypted vaults help manage data and results, supporting fields like bioinformatics, physics simulations, or social sciences with scalable compute.
IT teams can integrate GPU CLI into CI/CD pipelines for automated ML training, testing, and deployment. Using commands for serverless endpoints, daemon control, and organization management, it facilitates multi-user collaboration, cost tracking, and secure execution across projects, enhancing efficiency in industries like finance or healthcare for predictive analytics.
Offer tiered subscriptions for access to GPU resources, with pricing based on GPU type, usage time, and features like volumes or serverless endpoints. Revenue comes from monthly or annual fees, targeting individual developers, startups, and enterprises needing scalable compute without upfront hardware costs, with upsells for premium support or higher-tier GPUs.
Charge users based on actual GPU usage, measured in pod hours or compute minutes, with transparent pricing shown via commands like 'gpu inventory'. This model appeals to sporadic users or projects with variable demands, generating revenue from on-demand consumption, and can include additional fees for data transfer, storage, or premium templates.
Sell enterprise licenses with advanced features like organization management, service accounts, and dedicated support. Revenue is generated through annual contracts, customization services, and training, targeting large companies in sectors like finance or healthcare that require compliance, security, and integration with existing systems for AI workloads.
💬 Integration Tip
Integrate GPU CLI into existing workflows by using its JSON output options for automation, and leverage volume sync for persistent data across runs to reduce setup time.
Remote-control tmux sessions for interactive CLIs by sending keystrokes and scraping pane output.
Command-line tool to manage Google Workspace services including Gmail, Calendar, Drive, Sheets, Docs, Slides, Contacts, Tasks, People, Groups, and Keep.
Runs shell commands inside a dedicated tmux session named claw, captures, and returns the output, with safety checks for destructive commands.
A modern text-based browser. Renders web pages in the terminal using headless Firefox.
Write robust, portable shell scripts. Use when parsing arguments, handling errors properly, writing POSIX-compatible scripts, managing temp files, running commands in parallel, managing background processes, or adding --help to scripts.
NotebookLM CLI wrapper via `node {baseDir}/scripts/notebooklm.mjs`. Use for auth, notebooks, chat, sources, notes, sharing, research, and artifact generation/download.