Agent Skills for discovering GPU compute, comparing task fit, requesting an explicit quote approval, running bounded work and exporting results. For developers outside mainland China only; see https://cloudgpu.app/terms.
cloudgpu-discover-compute: discover live GPU capabilities and compare workload requirements.cloudgpu-lora-finetune: plan approved LoRA / QLoRA training and retrieve checkpoints.cloudgpu-comfyui-images: prepare reviewed ComfyUI workflows and export images.cloudgpu-ai-short-drama: compare short-drama production approaches, plan a 30-second storyboard, and price open-model video stages with approval gates.cloudgpu-open-model-inference: plan GPU inference and evaluation with protected inputs and verified outputs.
With the open Agent Skills CLI:
npx skills add cloudgpu-app/cloudgpu-agent-skillsSelect only the skills and clients you need. Installing a skill does not rent a GPU or approve spending.
Try the anonymous Compute Lab · Live readiness · Verified small task recipes
Download this repository's source archive from GitHub and extract it. From the extracted repository directory, copy a selected skill into your client's skills directory:
mkdir -p ~/.agents/skills
cp -R cloudgpu-discover-compute ~/.agents/skills/Repeat for other skills you need. Follow your client's current Agent Skills loading instructions and reload if required. Format compatibility is not a claim that every client has been tested.
Client-by-client setup · Quote → approve → run → save → release
- Read https://cloudgpu.app/docs/agents and https://cloudgpu.app/openapi.json for the current contract.
- Discover live GPU capabilities and request a quote before paid work.
- Present workload, data permissions, cost estimate and hard time/spend limits for explicit user approval. Connecting an account does not approve a job.
- Supply a reviewed workload command, pinned dependencies, compatible licensed models, authorized inputs and acceptance checks before execution.
- Download and verify outputs before destroying an interactive instance, or explicitly save them to Workspace / configured Job output storage and verify the saved result. Instance disks are erased on termination. Retention, funded-account access, storage budgets and output limits are separate from GPU rental; check the current contract before promising persistence.
The skills describe documented interfaces, not a guarantee that arbitrary workflows have been tested end to end. The linked small internal task recipes have published execution evidence; those results do not validate every skill, model or workflow. Keep credentials in a secure client credential mechanism; never include secrets in chat, source or logs.
REST documentation: https://cloudgpu.app/docs/agents
MCP documentation: https://cloudgpu.app/docs/mcp
The MCP server is available as a separate local stdio package:
npx -y cloudgpu-mcpFor remote clients, https://cloudgpu.app/mcp is the Streamable HTTP endpoint with OAuth. Follow the current MCP documentation for authorization and the transport-specific tool list. For local stdio, use connect_account and finish_connect or your client's secure credential mechanism. These skills do not bundle the MCP server; connecting does not approve paid work.
GPU rental uses source-specific billing: minute-billed sources round partial minutes up; whole-hour sources have a one-hour minimum and charge each started hour in full, including an early voluntary stop. Obtain a current quote, show its minimum charge, billing increment and start condition, and obtain explicit approval before allocation. Do not infer billing from a GPU model or hourly rate. Storage has separate budgets, usage charges and retention terms; consult the current API contract.
MIT. See LICENSE. Model, dataset and third-party workflow licenses remain separate.
- Compute Lab: anonymous live-check link, interactive billing illustration and historical task evidence.
- CrewAI readiness example: an actual framework tool with no LLM call or GPU allocation.
- Distribution source: static Hugging Face Space, technical DEV article and platform-specific publishing constraints.
The demo adds no application analytics. Hosting providers may collect operational telemetry under their own policies; external services linked from the demo have their own policies.