@alvin0/ai-agent-sdk-mcp

Runtime: Universal (Edge/Worker, browser, Deno, Bun, and Node).

pnpm add @alvin0/ai-agent-sdk-core @alvin0/ai-agent-sdk-mcp

Universal MCP client for Fetch-shaped HTTP runtimes. It connects remote MCP servers to the SDK as a versioned ToolSource.

import { createAgentRuntime } from '@alvin0/ai-agent-sdk-core'
import { connectMcpHttp } from '@alvin0/ai-agent-sdk-mcp'

const runtime = await createAgentRuntime({ providers: [modelProvider] })
let connection: Awaited<ReturnType<typeof connectMcpHttp>> | undefined
try {
  connection = await connectMcpHttp({
    serverName: 'tools',
    url: 'https://tools.example.com/mcp',
    logger: runtime.logger({ fields: { integration: 'mcp-http' } }),
  })
  const agent = runtime.agent({ model, toolSources: [connection] })
  await agent.generate('Use the connected tools when needed.')
} finally {
  try {
    await runtime.close()
  } finally {
    await connection?.closeWithReport()
  }
}

The normal root and /client route contain no server, stdio, filesystem, node:http, or process lifecycle integration. Use @alvin0/ai-agent-sdk-mcp-node for the Node stdio client, and a dedicated server package for hosting.

HTTP endpoints are host-selected. HTTPS, public-network-only access, and no redirects are the defaults; local development must opt out explicitly. Use allowedOrigins, the async validateEndpoint hook, a DNS-pinning custom fetch or network egress policy, response/catalog/result bounds, and operation deadlines according to the application's trust boundary. OAuth credential persistence and redirect handling remain caller-owned.

Composition: runtime-agent.toolSources. Lifecycle: connected-caller-owned; create the runtime first, pass runtime.logger(...) while connecting, close the runtime to quiesce runs, then inspect connection.closeWithReport().