Skip to main content

​
Building a chat loop

With mcp-use you can build interactive interface where users can have conversations with your MCPAgent, maintaining context and memory across multiple queries.

​
Basic chat loop

Here’s a basic chat-loop with conversation memory enabled: 
import readline from 'node:readline'
import { config } from 'dotenv'
import { ChatOpenAI } from '@langchain/openai'
import { MCPAgent, MCPClient } from 'mcp-use'

config() // Load environment variables

async function basicChatLoop() {
    // MCP server configuration
    const configuration = {
        mcpServers: {
            playwright: {
                command: 'npx',
                args: ['@playwright/mcp@latest'],
                env: { DISPLAY: ':1' }
            },
            filesystem: {
                command: 'npx',
                args: ['-y', '@modelcontextprotocol/server-filesystem', '/tmp']
            }
        }
    }

    // Create client and agent
    const client = new MCPClient(configuration)
    const llm = new ChatOpenAI({ model: 'gpt-4o' })

    const agent = new MCPAgent({
        llm,
        client,
        memoryEnabled: true, // Enable memory to track conversation history
        maxSteps: 20
    })

    // Some initial messages
    console.log('🤖 MCP Agent Chat')
    console.log('Type \'quit/exit\' to exit the chat.')
    console.log('Type \'clear\' to clear conversation history')

    // Create readline interface
    const rl = readline.createInterface({
        input: process.stdin,
        output: process.stdout
    })

    const question = (prompt: string): Promise<string> => {
        return new Promise((resolve) => {
            rl.question(prompt, resolve)
        })
    }

    try {
        while (true) {
            const userInput = await question('\nYou: ')

            if (['quit', 'exit'].includes(userInput.toLowerCase())) {
                console.log('👋 Goodbye!')
                break
            }

            if (userInput.toLowerCase() === 'clear') {
                agent.clearConversationHistory()
                console.log('🧹 Conversation history cleared.')
                continue
            }

            // Skip empty messages
            if (!userInput) {
                continue
            }

            try {
                process.stdout.write('\n🤖 Assistant: ')
                const response = await agent.run(userInput)
                console.log(response)
            } catch (error) {
                console.error(`\n❌ Error: ${error}`)
                console.log('Please try again or type \'exit\' to quit.')
            }
        }
    } finally {
        rl.close()
        await client.closeAllSessions()
    }
}

basicChatLoop().catch(console.error)

​
Streaming Chat Loop

Here’s a chat loop with streaming responses enabled: 
import readline from 'node:readline'
import { config } from 'dotenv'
import { ChatOpenAI } from '@langchain/openai'
import { MCPAgent, MCPClient } from 'mcp-use'

config() // Load environment variables

async function streamingChatLoop() {
    // MCP server configuration
    const configuration = {
        mcpServers: {
            playwright: {
                command: 'npx',
                args: ['@playwright/mcp@latest'],
                env: { DISPLAY: ':1' }
            }
        }
    }

    // Create client and agent
    const client = new MCPClient(configuration)
    const llm = new ChatOpenAI({ model: 'gpt-4o' })

    const agent = new MCPAgent({
        llm,
        client,
        memoryEnabled: true, // Enable memory to track conversation history
        maxSteps: 20
    })

    // Some initial messages
    console.log('🤖 MCP Agent Chat (Streaming)')
    console.log('Type \'quit/exit\' to exit the chat.')
    console.log('Type \'clear\' to clear conversation history')

    const rl = readline.createInterface({
        input: process.stdin,
        output: process.stdout
    })

    const question = (prompt: string): Promise<string> => {
        return new Promise((resolve) => {
            rl.question(prompt, resolve)
        })
    }

    try {
        while (true) {
            const userInput = await question('\nYou: ')

            if (['quit', 'exit'].includes(userInput.toLowerCase())) {
                console.log('👋 Goodbye!')
                break
            }

            if (userInput.toLowerCase() === 'clear') {
                agent.clearConversationHistory()
                console.log('🧹 Conversation history cleared.')
                continue
            }

            if (!userInput) { // Skip empty messages
                continue
            }

            try {
                process.stdout.write('\n🤖 Assistant: ')

                // Stream the response
                for await (const step of agent.stream(userInput)) {
                    // In TypeScript, stream returns steps rather than text chunks
                    // You might want to use streamEvents for token-level streaming
                    console.log(`\nTool: ${step.action.tool}`)
                    console.log(`Result: ${step.observation}`)
                }
            } catch (error) {
                console.error(`\n❌ Error: ${error}`)
                console.log('Please try again or type \'exit\' to quit.')
            }
        }
    } finally {
        rl.close()
        await client.closeAllSessions()
    }
}

streamingChatLoop().catch(console.error)

​
Chat Loop with Structured I/O

It’s possible to create a chat loop that can handle both natural language and structured inputs, allowing users to request specific tasks or analyses in a structured format. Here’s an example of how to implement this: 
import readline from 'node:readline'
import { z } from 'zod'
import { config } from 'dotenv'
import { ChatOpenAI } from '@langchain/openai'
import { MCPAgent, MCPClient } from 'mcp-use'

config() // Load environment variables

// Define the schema using Zod
const TaskRequest = z.object({
    taskType: z.string().optional().describe('The type of task to perform'),
    description: z.string().optional().describe('Detailed description of the task'),
    priority: z.string().optional().describe('Priority level: low, medium, high')
})

type TaskRequest = z.infer<typeof TaskRequest>

async function structuredChatLoop() {
    // MCP server configuration
    const configuration = {
        mcpServers: {
            playwright: {
                command: 'npx',
                args: ['@playwright/mcp@latest'],
                env: { DISPLAY: ':1' }
            }
        }
    }

    // Create client and agent
    const client = new MCPClient(configuration)
    const llm = new ChatOpenAI({ model: 'gpt-4o' })

    const agent = new MCPAgent({
        llm,
        client,
        memoryEnabled: true, // Enable memory to track conversation history
        maxSteps: 20
    })

    // Initial messages
    console.log('🤖 MCP Agent Chat (Structured)')
    console.log('You can chat naturally or request structured task analysis')
    console.log('Type \'task\' to create a structured task request')

    const rl = readline.createInterface({
        input: process.stdin,
        output: process.stdout
    })

    const question = (prompt: string): Promise<string> => {
        return new Promise((resolve) => {
            rl.question(prompt, resolve)
        })
    }

    try {
        while (true) {
            const userInput = await question('\nYou: ')
            if (['exit', 'quit'].includes(userInput.toLowerCase())) {
                console.log('👋 Goodbye!')
                break
            }

            try {
                if (userInput.toLowerCase() === 'task') {
                    console.log('\n📋 Creating structured task...')
                    const taskDescription = await question('Describe your task: ')

                    const task = await agent.run(
                        `Analyze a task with the following description: ${taskDescription}`,
                        undefined, // maxSteps
                        undefined, // manageConnector
                        undefined, // externalHistory
                        TaskRequest // output schema
                    )

                    // Print task analysis
                    console.log('\n✅ Task Analysis:')
                    console.log(`• Type: ${task.taskType}`)
                    console.log(`• Description: ${task.description}`)
                    console.log(`• Priority: ${task.priority || 'low'}`)

                    const proceed = await question('\nDo you want to proceed with this task? (y/n)')
                    if (proceed.toLowerCase() === 'y') {
                        const response = await agent.run(
                            `Execute the following task: ${task.description}`
                        )
                        console.log(`\n🤖 Assistant: ${response}`)
                    }
                } else {
                    // Regular conversation
                    const response = await agent.run(userInput)
                    console.log(`\n🤖 Assistant: ${response}`)
                }
            } catch (error) {
                console.error(`❌ Error: ${error}`)
                console.log('Please try again or type \'exit\' to quit.')
            }
        }
    } finally {
        rl.close()
        await client.closeAllSessions()
    }
}

structuredChatLoop().catch(console.error)

​
Next Steps

Agent Configuration

Learn more about configuring agents for optimal streaming performance

Multi-Server Setup

Stream output from agents using multiple MCP servers

Security Best Practices

Learn how to secure your MCP deployments