AI Agent with Notion via MCP

Scenario

Developers use this workflow when they need a Qwen-powered agent on Alibaba Cloud Bailian to autonomously query, update, and organize Notion workspaces using natural language. By exposing Notion as an MCP tool, the agent bypasses rigid UI constraints and performs structured knowledge management through dynamic tool calling.

Integration steps

1. Create Notion Internal Integration: In Notion Settings → Integrations, create a new integration and copy the NOTION_TOKEN (format: ntn_xxxxxxxxxxxx).
2. Share Workspace Content: Open target Notion pages/databases, click ••• → Add connections, and select your integration to grant read_content and insert_content scopes.
3. Initialize MCP Server: Scaffold with npm create @modelcontextprotocol/server@latest notion-mcp and install @notionhq/client. Set environment variables:

   export NOTION_API_KEY="ntn_xxxxxxxxxxxx"
   export MCP_PORT=3001

1. Define MCP Tools: Register tools that map to Notion API v1 endpoints:

   server.tool("notion_search", { query: z.string() }, async ({ query }) => {
     const res = await notion.search({ query, filter: { property: "object", value: "page" } });
     return { content: [{ type: "text", text: JSON.stringify(res.results) }] };
   });

1. Start MCP Server: Run node index.js. Verify connectivity with curl -X POST http://localhost:3001/mcp -H "Content-Type: application/json" -d '{"jsonrpc":"2.0","method":"tools/list","id":1}'.
2. Configure Bailian Agent: In Bailian Console → Agent Builder → Tools, add an MCP endpoint. Set server_url to http://<host>:3001/mcp and transport to sse. Enable auto_tool_selection.
3. Bind & Test: In the Bailian SDK, initialize with tools: ["notion_search", "notion_create_page"]. Send: "Find all Q3 OKRs in Notion and summarize them."

Architecture

The user prompt hits the Bailian orchestration layer, where Qwen evaluates registered MCP tool definitions. When a Notion-related intent is detected, Bailian’s MCP client sends a JSON-RPC request over SSE to the hosted MCP server. The server translates the call into authenticated POST /v1/search or POST /v1/pages requests against Notion’s REST API, returning structured JSON. Bailian parses the response, injects it into Qwen’s context window for reasoning, and generates the final output. Notion acts purely as the data layer; Bailian handles prompt routing, context management, and tool execution.

Prerequisites

• Alibaba Cloud Bailian account with Qwen API access enabled
• Notion workspace with Admin/Editor permissions
• Node.js v18+ and npm
• @notionhq/client and @modelcontextprotocol/sdk packages
• Publicly accessible endpoint or tunnel (e.g., ngrok) for remote MCP hosting

Common pitfalls

• Missing Notion scopes: The integration token lacks read_content or insert_content, causing 403 Forbidden on MCP calls. Verify scopes in Notion Developer Console.
• MCP transport mismatch: Bailian expects sse but the server defaults to stdio; explicitly configure transport: "sse" in your MCP server initialization.
• Tool schema drift: Overly complex Zod schemas cause Qwen to generate malformed JSON arguments. Keep parameters flat, required, and explicitly typed.
• Notion rate limits: Rapid search invocations trigger 429 Too Many Requests. Implement client-side caching or debounce tool calls in your MCP server.

Typical questions

• build notion ai agent with bailian
• connect llm to notion via mcp
• notion mcp tool integration
• ai agent read write notion
• bailian agent notion integration
• 用百炼大模型对接Notion
• MCP集成Notion工具
• AI智能体操作Notion数据库

Q: How do I build an AI agent that connects to Notion via MCP using Alibaba Cloud Bailian? A: You can integrate a Bailian AI agent with Notion by deploying a custom MCP server that exposes Notion API endpoints and configuring the Bailian Agent Builder to call it via SSE. First, create a Notion internal integration with read and insert scopes, then scaffold the MCP server using the provided npm package and set the required environment variables. Finally, add the MCP endpoint to your Bailian agent's tools, enable auto tool selection, and bind the specific tool names to test natural language queries.