//! luxe-mcp — Model Context Protocol orchestration. //! //! The problem this solves: MCP tool definitions are verbose. Registering every tool from every //! server up front burns thousands of context tokens whether or not the agent uses them, and eager //! connections spawn processes at launch. Instead Luxe exposes **one** proxy tool ([`McpProxyTool`], //! `mcp`) — the agent *discovers* tools on demand (`search` → `describe` → `call`) — backed by a //! [`ServerRegistry`] that connects servers **lazily** and caches their tool metadata so discovery //! works without a live connection. //! //! Architecture: the low-level stdio transport + MCP types live in `luxe-core` (no HTTP dep //! there); this crate carries the orchestration plus the Streamable-HTTP transport and OAuth that //! core's no-HTTP rule keeps out of core. MCP servers stay untrusted — calls are permission-gated //! and results are surfaced as data, never instructions. //! //! Both transports are reached through [`connect`], so nothing has to know which one a server //! uses. A remote server is configured with a URL and, where it needs one, a bearer token or an //! OAuth flow — which is how a vendor's hosted MCP server (Firecrawl, Exa, and the rest) is used //! from Luxe without Luxe carrying an adapter for it. pub mod http; pub mod oauth; mod registry; mod secret; pub mod token_store; mod tool; pub use http::McpHttpError; pub use oauth::OAuthError; pub use registry::{open_connection as connect, RegistryError, ServerRegistry, ToolRef}; pub use secret::SecretError; pub use token_store::TokenStoreError; pub use tool::McpProxyTool;