Claude Skills but Local (and Private) (instavm.io)

Anthropic’s new “Skills” feature has been integrated into Coderunner, an open‑source local sandbox, allowing Skills to run against local LLMs and files without needing the Claude platform. That means developers can import existing Claude Skills unchanged, run them locally (Python/JS), and keep data private — a major step for local-first AI workflows and agent development. The team demoed the flow with an ascii‑art skill via MCP (Modular Connector Protocol) calls, showing Skills discovery, metadata retrieval, and execution all happening inside Coderunner. Technically, Coderunner preserves the Claude skills directory layout so Skills work out‑of‑the‑box, but it solves the lack of a mounted host filesystem (Claude mounts /mnt) by exposing multiple MCP endpoints: one for frontmatter, one for SKILL.md, and one for execution, enabling progressive loading of only needed content. Responses are rewritten on the fly to map Claude’s /mnt/user-data to Coderunner’s /app/uploads, and local assets are exposed through ~/.coderunner/assets/outputs. Because Coderunner supports standard MCP, other tools (e.g., Gemini CLI) can use the same Skills, giving developers more freedom (no enforced internet/library restrictions) and stronger privacy guarantees while retaining interoperability with existing Skill ecosystems.