The emerging standard for how AI models talk to tools. Anthropic shipped it; everyone is adopting it. Here is what it is.
Before MCP, every AI host had its own tool-calling format and every tool had to be re-implemented for each host. The result was massive duplication and brittle integrations. MCP standardises (1) tool discovery, (2) tool invocation, (3) prompt resources, and (4) sampling — so a tool written once works everywhere.
undefined
Anthropic (Claude Desktop, Claude Code) ships first-class MCP support. Cursor, Zed, Windsurf, Continue, and many code editors adopted MCP through 2025-2026. Many third-party tools — GitHub, Linear, Notion, Postgres, browser automation — now ship official MCP servers. Luna supports MCP servers as one of her tool-discovery paths.
MCP runs locally by default (stdio transport), which is great for security but limits some patterns. The HTTP transport is improving but still less mature. Authentication is the rough edge — multi-user remote MCP servers are an active area of work in 2026.
Luna ships native MCP support — she can connect to MCP servers you point her at (GitHub, your filesystem, Postgres, a custom tool you wrote) and use them inside her agent loop. The MCP layer is one of 92+ ways Luna picks up capabilities.
You do not need to know MCP exists to use Luna. But if you live in MCP-land already, your stack will feel at home.
Connect an MCP server to Luna →
No. MCP was created by Anthropic but is an open protocol. OpenAI, Google, sovereign systems (including Luna) and most major open-source LLM frameworks now speak MCP.
The protocol itself is fine. The security question is "what is the MCP server allowed to do?" If a server can run shell commands, it can do anything you can. Treat MCP servers like browser extensions — install only what you trust, and read the manifest.
Yes, easily. Anthropic publishes SDKs in TypeScript, Python and others. A useful MCP server (e.g. exposing a custom internal API) is usually under 200 lines of code.
Function calling is a feature inside a single LLM API (you pass tool definitions in each request). MCP is a protocol that lives outside the LLM — any MCP-compatible host can use any MCP server without coordinating per-app. Function calling is local; MCP is the network.