Server Overview

The McpServer class exposes capabilities (tools, resources, prompts) to connected MCP clients.

Creating a Server

auto transport = std::make_shared<StdioServerTransport>();

ServerOptions opts;
opts.server_info = Implementation{"MyServer", "1.0.0"};
opts.capabilities = ServerCapabilities{};
opts.capabilities->tools = ToolsCapability{};

auto server = McpServer::Create(std::move(transport), opts);
server->Run();

For Streamable HTTP mode:

StreamableHttpServerOptions http_opts;
http_opts.port = 3001;
http_opts.stateless = true;  // or false for session mode

auto transport = std::make_shared<StreamableHttpServerTransport>(http_opts);

ServerOptions opts;
opts.server_info = Implementation{"MyServer", "1.0.0"};
opts.capabilities = ServerCapabilities{};
opts.capabilities->tools = ToolsCapability{};

auto server = McpServer::Create(std::move(transport), opts);
server->Run();

ServerOptions

Field	Type	Description
server_info	optional<Implementation>	Server identity (name, version)
capabilities	optional<ServerCapabilities>	Declared capabilities
protocol_version	optional<string>	Pin to a specific version
server_instructions	optional<string>	Instructions sent to client
initialization_timeout	chrono::seconds	Handshake timeout (default 60s)
validate_tool_input	bool	Enable JSON Schema input validation
validate_tool_output	bool	Enable JSON Schema output validation
task_store	shared_ptr<IMcpTaskStore>	Task persistence backend
request_state_verifier	function<bool(string_view)>	HMAC/AEAD verifier for MRTR
cache_hints	optional<map<string, CacheHint>>	Per-method cache hints (ttlMs, cacheScope)
input_required_config	optional<InputRequiredConfig>	Configuration for MRTR/elicitation behavior
input_required_config.max_rounds	int	Maximum elicitation rounds (default 8)
input_required_config.round_timeout	chrono::seconds	Per-round timeout (default 600s)
input_required_config.legacy_shim	bool	Enable legacy compatibility shim (default true)
on_request	function	Called on each incoming JSON-RPC request with method name + full request body
on_response	function	Called on each outgoing JSON-RPC response
on_error	function	Called on each outgoing JSON-RPC error response
on_notification	function	Called on each incoming JSON-RPC notification
on_method_called	function(string_view)	Shorthand — method name only (fires alongside on_request)
on_protocol_error	function(string_view)	Shorthand — error message only (fires alongside on_error)
on_client_connected	function(Implementation)	Called when a client completes initialize
on_initialized	function()	Called when client sends notifications/initialized
on_transport_close	function()	Called when the transport connection closes
on_transport_error	function(string_view)	Called on transport-level errors
incoming_filters	shared_ptr<FilterPipeline>	Pipeline to intercept/modify/block inbound messages
outgoing_filters	shared_ptr<FilterPipeline>	Pipeline to intercept/modify/block outbound messages

Shorthand callbacks (on_method_called, on_protocol_error) and full-message callbacks (on_request, on_error) are chained — both fire when set simultaneously.

Lifecycle

1. Construction: McpServer::Create — creates the session handler, wires handlers, starts message loop
2. Registration: Register tools, resources, prompts, extensions
3. Run: server->Run() — blocks on io_context.run()
4. Shutdown: server->Close() — cancels pending requests, closes transport

Capability Derivation

Capabilities are automatically derived from registered primitives. For example, registering a tool sets capabilities.tools.list_changed = true. The ServerOptions.capabilities field is declared but currently not consumed by the server — capabilities are always auto-derived from registered tools, resources, prompts, and task store.