Monorch

MCP

Connect Model Context Protocol servers over stdio or Streamable HTTP (with SSE fallback), then register their tools as normal tool() definitions.

Stdio transport

Spawns a local MCP server process. Use this for filesystem / desktop-style servers.

mcp-stdio.ts
import { mcpStdio, mcpTools, agent } from "@monorch/ai";

const session = await mcpStdio({
  command: "npx",
  args: ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
});

const tools = await mcpTools(session, {
  prefix: "fs_",
  permission: { type: "roles", roles: ["agent"] },
});

const bot = agent({
  model,
  tools,
  instructions: "Use filesystem tools when needed.",
});

// later
await session.close();

HTTP transport

Connects to a remote MCP endpoint. Default mode tries Streamable HTTP, then falls back to legacy SSE.

mcp-http.ts
import { mcpHttp, mcpTools } from "@monorch/ai";

const session = await mcpHttp({
  url: "https://mcp.example.com/mcp",
  headers: { authorization: `Bearer ${token}` },
  // transport: "streamable-http" | "sse" | "auto" (default)
});

const tools = await mcpTools(session, { prefix: "remote_" });
await session.close();

Mock transport (tests)

mcp-mock.ts
import { mcpTools, mockMcp } from "@monorch/ai";

const remote = mockMcp([
  {
    name: "lookup_order",
    description: "Look up an order",
    inputSchema: {
      type: "object",
      properties: { orderId: { type: "string" } },
      required: ["orderId"],
    },
    execute: (args) => ({ orderId: args.orderId, status: "paid" }),
  },
]);

await mcpTools(remote, { prefix: "mcp_" });
// registers mcp_lookup_order with JSON Schema → Rust IR

Transport shape

transport.ts
type McpTransport = {
  listTools(): Promise<McpToolInfo[]> | McpToolInfo[];
  callTool(name: string, args: unknown): Promise<JsonValue> | JsonValue;
};

type McpSession = McpTransport & { close(): Promise<void> };

BYO transports still work — implement list/call yourself if you already wrap another SDK.

FAQ