Skrun converts any agent skill defined in a SKILL.md file into a stateful POST /run API endpoint with multi-model support.

Primary GitHub documentation shows Skrun uses agent.yaml for typed inputs/outputs, permissions, and tests while supporting Anthropic, OpenAI, Google, Mistral, and Groq models with automatic fallback and key-value state persistence across calls (https://github.com/skrun-dev/skrun). Related LangGraph patterns for persistent agent memory and the ReAct framework's emphasis on tool-calling loops are synthesized here through both local CLI scripts and MCP servers compatible with Claude Desktop (https://python.langchain.com/docs/langgraph/; https://arxiv.org/abs/2210.03629).

Mainstream coverage of agentic AI has focused on model releases and reasoning improvements but missed the deployment friction of turning one-off skills from Copilot or Claude Code into callable production services, a gap Skrun addresses by allowing skrun init --from-skill to import existing SKILL.md files and generate configs in seconds. Original source accurately lists CLI commands (init, dev, test, deploy, push) and demo agents such as pdf-processing and seo-audit but does not connect this standardization to the emerging agent skill marketplace dynamic visible in 2024 enterprise pilots.

The local runtime and RuntimeAdapter interface prepare for cloud deployment, enabling any imported skill to become a versioned, authenticated API that maintains state on repeated calls as demonstrated in the seo-audit example.