Magic is moving toward a small semantic execution core that can be embedded by host software, inspected by humans, and consumed by maintenance agents.

Semantic Catalog

  • document every built-in semantic function with aliases, argument forms, return type, side effects, and examples;
  • separate core semantics from host-provided semantics such as MagicPaper and MagicRedis;
  • export the catalog from runtime state instead of relying only on manually maintained text.

Concurrency Safety

  • define which values are local to one SpellContext and which values are shared through a host ContextMap;
  • document safe patterns for async execution, daemon async execution, HTTP requests, and MCP sessions;
  • make env/session isolation explicit so projects, servers, and agents do not leak state into each other.

Agent-Readable Catalog

  • expose semantic metadata as JSON;
  • include host capability catalogs, starting with MagicPaper;
  • make magic mcp able to answer catalog queries without scraping README text;
  • provide examples that map user intent to compact Magic scripts.