iFlow Search MCP server

iflow-search-js

JavaScript / TypeScript integrations for iFlow Search (心流搜索) — a search API that provides web search, image search, and web page fetching with AI-friendly structured output.

• Product:
• API docs:
• Operator: 杭州星辰千寻科技有限公司

This monorepo contains:

• A framework-agnostic core SDK
• A LangChain JS tool adapter (also usable from LangGraph)
• A LangGraph agent example
• An MCP stdio server for Hermes / Claude Code / Claude Desktop / other MCP clients

Package / usage matrix

Why there is no @iflow-ai/search-langgraph

LangGraph consumes LangChain tools directly. A separate search-langgraph package would be a thin re-export with no added value — use @iflow-ai/search-langchain for both. See examples/langgraph-agent for a working createReactAgent wiring.

Current status

• ✅ @iflow-ai/search-core — implemented, framework-agnostic client
• ✅ @iflow-ai/search-langchain — implemented, three tools (iflowwebsearch, iflowimagesearch, iflowwebfetch)
• ✅ examples/langgraph-agent — implemented, ReAct agent end-to-end smoke validated against real iFlow API
• ❌ no separate @iflow-ai/search-langgraph package (intentional — see above)
• ✅ @iflow-ai/search-mcp — implemented, stdio MCP server with three tools mirroring the LangChain adapter; optional MCP-host attribution via IFLOWMCPCLIENT / IFLOWMCPCLIENT_VERSION
• ✅ plugins/iflow-search — Claude Code Plugin manifest (metadata-only, runs npx -y @iflow-ai/search-mcp); awaiting submission to the Anthropic Plugin Directory. See plugins/iflow-search/README.md.

Workspace development

pnpm install
pnpm -r run typecheck
pnpm -r run build
pnpm -r run test

The workspace pins @langchain/core to ^1.1.44 via pnpm-workspace.yaml overrides: to keep LangGraph's runtime and LangChain's tool types on a single major. Removing this override re-introduces ToolMessage cross-version serialization bugs in LangGraph agent loops; rerun the LangGraph agent tests before you touch it.

Basic usage — core SDK

import { createIFlowSearchClient } from "@iflow-ai/search-core";

const client = createIFlowSearchClient({
  apiKey: process.env.IFLOW_API_KEY!,
  source: "core",
  integrationName: "my-app",
  integrationVersion: "1.0.0",
});

const result = await client.webSearch({ query: "flash attention", count: 5 });
if (!result.ok) {
  console.error(result.error.code, result.error.message);
} else {
  for (const r of result.data.results) console.log(r.title, r.url);
}

Shortened. The full README is on GitHub.