Model Context Protocol (MCP)

An MCP server's tools reach an agent the same way a @tool does, once the server is named in .mcp.json and the agent asks for it by name.

{
  "mcpServers": {
    "search": {
      "type": "http",
      "url": "http://localhost:8765/mcp"
    }
  }
}

from agentdeck import Agent, Deck

researcher = Agent(
    name="Researcher",
    instructions="Use the search tools to answer questions about current events.",
    mcp=["search"],
)

deck = Deck(agents=[researcher], mcp=".mcp.json")

mcp=[...] names servers, resolved against the Deck's own .mcp.json at build() - the same pattern as skills=[...]. Deck.from_project() finds .mcp.json on its own, next to .agentdeck/ (a sibling, not inside it).

HTTP only, today. type accepts http (or streamable-http); no stdio transport is wired up. headers and timeout are optional per-server.

When a server is down

A server that fails to connect at startup does not fail the boot: it is marked unavailable, and an agent naming it gets a banner in its own instructions instead of the tools:

## Runtime MCP status

The following MCP server(s) you depend on are UNAVAILABLE: `search`.
Tools from those servers are NOT exposed to you in this turn.
If the user's request requires those tools, reply with exactly:

    error: mcp_unavailable: <server-name>

and stop. Do not invent results. Do not retry. Do not call other tools as a substitute unless
the user explicitly asks for an alternative.

The banner is prepended to the agent's instructions; the model fills in <server-name> itself rather than guessing or substituting another tool. An agent naming an MCP server the Deck was never given mcp= for, or a name outside its own .mcp.json, fails at build() instead - that is a wiring mistake, not a runtime outage.