Mead.MCP.Policy (mead v0.1.1)

Copy Markdown View Source

Decides which tools the model sees and whether a named call may run. Pure: it reads no process and asks no server.

The inputs are the catalog entries (what the connected servers offer), the normalized servers of Mead.MCP.Configuration (what the operator allowed, with a class per tool), and the options of the turn. A tool is visible only if the server offers it, the configuration lists it, and its class is allowed for this turn: :read always, :effect only with tools: :effect. Any other value of tools: is the default, because an unclear request must not widen access.

What a server says about itself, such as an annotation that a tool is read-only, is never consulted. The model sees server__tool; the first __ splits it, which is unambiguous because a configured server name has no underscore.

Summary

Types

class()

Functions

authorize(model_name, entries, servers, options)
model_name_valid?(server_name, tool)

Whether server__tool is a name the model can be shown. A tool whose name fails this is never visible, so the configuration refuses it at boot instead of hiding it without a word.

visible_tools(entries, servers, options)

Types

class()

@type class() :: :read | :effect

Functions

authorize(model_name, entries, servers, options)

@spec authorize(term(), [map()], [map()], keyword()) ::
  {:ok, %{server: atom(), tool: String.t(), class: class()}}
  | {:error, :unknown_tool | :not_allowed}

model_name_valid?(server_name, tool)

@spec model_name_valid?(atom(), String.t()) :: boolean()

Whether server__tool is a name the model can be shown. A tool whose name fails this is never visible, so the configuration refuses it at boot instead of hiding it without a word.

visible_tools(entries, servers, options)

@spec visible_tools([map()], [map()], keyword()) :: [map()]