The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes standard package managers and container registries to download well-known, reputable software components.
Installs the vllm package via pip.
Pulls official Docker images from vllm/vllm-openai.
[INDIRECT_PROMPT_INJECTION]: The skill describes how to deploy a server that ingests untrusted data for LLM inference, creating a surface for indirect prompt injection.
Ingestion points: User prompts are accepted via the /v1/chat/completions API endpoint.
Boundary markers: Includes a "Version and exposure boundary" section that instructs users to use loopback-only addresses (127.0.0.1) and warns against using sample keys.
Capability inventory: The skill utilizes subprocess execution for model serving (vllm serve) and container management (docker run).
Sanitization: Documentation explicitly recommends implementing approved authentication, TLS, and network policies before exposing the service remotely.
[SAFE]: The skill uses placeholders such as your-secret-key, ${HF_TOKEN}, and ${VLLM_API_KEY} instead of hardcoding sensitive credentials.

vllm-server