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: plan open-model video stages with a project budget and review gates.cloudgpu-open-model-inference: plan GPU inference and evaluation with protected inputs and verified outputs.
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.
- 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. A batch command must export results to user-authorized storage before exit. Platform disks are erased on termination; native retained artifacts are not available.
The skills describe documented interfaces, not pretested end-to-end workload recipes. No paid GPU execution or model generation was performed to validate them. 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 a separate local stdio package:
npx -y cloudgpu-mcpConfigure it through your MCP client's secure credential mechanism. The website /mcp path is not a remote MCP endpoint. These skills do not bundle the MCP server.
Billed per minute, partial minutes count as a full minute, billing stops on destroy.
MIT. See LICENSE. Model, dataset and third-party workflow licenses remain separate.