Struct McpServerManager 

pub struct McpServerManager { /* private fields */ }

Manages multiple MCP server connections.

Implementations

impl McpServerManager

pub async fn start(configs: &HashMap<String, McpServerConfig>) -> Self

Start all configured MCP servers, initialize them, and discover tools.

Servers that fail to start are logged and skipped (non-fatal).

pub fn get_all_tools(&self) -> &[(String, McpToolDef)]

Get all discovered tools with their server names.

pub fn has_server(&self, name: &str) -> bool

True iff the named server started and has an active client, even if it advertised zero tools.

pub fn has_servers(&self) -> bool

Check if any MCP servers are active.

pub async fn call_tool( &self, server_name: &str, tool_name: &str, arguments: &Value, ) -> Result<McpToolResult>

Call a tool on a specific server.

Concurrency

Multiple concurrent calls to the same server will serialize at the transport layer (StdioTransport holds a mutex over stdin writes and uses a shared pending-response map for JSON-RPC correlation). This is intentional: JSON-RPC over stdio is a byte stream, and interleaved writes would corrupt messages. Calls to different servers run fully in parallel since each has its own transport.

pub fn format_tool_result( result: &McpToolResult, ) -> (String, Option<Vec<String>>)

Convert an MCP tool result into text suitable for a tool result message. Images are returned separately for multimodal attachment. Audio is attached through the same channel — adapters that don’t support audio will silently drop it. Resource links + embedded resources render as text so the model can follow up with another tool call.

pub async fn shutdown(&self)

Gracefully shut down all MCP servers.