• Creates a sandbox based on Bun 1.3
• Installs git and clones this repository
• Installs dependencies
• Configures the server to start automatically on port 3000

The build may take a few minutes. Once complete, your template is ready to use.

3. Use the Client Library

Install the client library in your project:

npm install @dzhng/claude-agent
# or
bun add @dzhng/claude-agent

Connect to your E2B sandbox:

import { ClaudeAgentClient } from '@dzhng/claude-agent'

const client = new ClaudeAgentClient({
  e2bApiKey: process.env.E2B_API_KEY,
  anthropicApiKey: process.env.ANTHROPIC_API_KEY,
  template: 'claude-agent-server', // Your E2B template name
  debug: true,
})

// Start the client (creates E2B sandbox and connects)
await client.start()

// Listen for messages from Claude
client.onMessage(message => {
  if (message.type === 'sdk_message') {
    console.log('Claude:', message.data)
  }
})

// Send a message to Claude
client.send({
  type: 'user_message',
  data: {
    type: 'user',
    session_id: 'my-session',
    message: {
      role: 'user',
      content: 'Hello, Claude!',
    },
  },
})

// Clean up when done
await client.stop()

That's it! The client library handles:

• Creating and managing E2B sandboxes
• WebSocket connection
• Message serialization
• Cleanup and resource management

Modifying the Server

If you want to customize the server behavior:

1. Edit Server Code

The server code is in packages/server/:

• index.ts - Main server and WebSocket handling
• message-handler.ts - Message processing logic
• const.ts - Configuration constants

2. Test Locally

Start the server locally:

bun run start:server

In another terminal, run the test client against localhost:

bun run test:local

This runs packages/client/example-client.ts connected to localhost:3000 instead of E2B.

3. Rebuild E2B Image

Once you're satisfied with your changes, rebuild the E2B template:

bun run build:e2b

Your updated server will be deployed to E2B with the same template name.

Available Scripts

bun run build:e2b

Builds and deploys the server as an E2B template. This is the main way to deploy your server to the cloud.

bun run test:client

Runs the example client (packages/client/example-client.ts) connected to an E2B sandbox. Requires both E2B_API_KEY and ANTHROPIC_API_KEY in your .env file.

bun run start:server

Starts the server locally on http://localhost:3000. Use this for local development and testing.

bun run test:local

Runs the example client connected to localhost:3000. Use this to test your local server changes before rebuilding the E2B image.

Client Library API

Installation

npm install @dzhng/claude-agent
# or
bun add @dzhng/claude-agent

Constructor Options

interface ClientOptions {
  // Required (unless using environment variables)
  anthropicApiKey?: string

  // E2B Configuration (optional if using connectionUrl)
  e2bApiKey?: string
  template?: string // E2B template name, defaults to 'claude-agent-server'
  timeoutMs?: number // Sandbox timeout, defaults to 5 minutes

  // Connection Configuration
  connectionUrl?: string // Use this to connect to local/custom server instead of E2B

  // Other Options
  debug?: boolean // Enable debug logging

  // Query Configuration (passed to server)
  agents?: Record<string, AgentDefinition>
  allowedTools?: string[]
  systemPrompt?:
    | string
    | { type: 'preset'; preset: 'claude_code'; append?: string }
  model?: string
}

Methods

• async start() - Initialize the client and connect to the server
• send(message: WSInputMessage) - Send a message to the agent
• onMessage(handler: (message: WSOutputMessage) => void) - Register a message handler (returns unsubscribe function)
• async stop() - Disconnect and clean up resources

Example: Connect to E2B

import { ClaudeAgentClient } from '@dzhng/claude-agent'

const client = new ClaudeAgentClient({
  e2bApiKey: process.env.E2B_API_KEY,
  anthropicApiKey: process.env.ANTHROPIC_API_KEY,
  template: 'claude-agent-server',
  debug: true,
})

await client.start()

client.onMessage(message => {
  if (message.type === 'sdk_message') {
    console.log('Claude:', message.data)
  }
})

client.send({
  type: 'user_message',
  data: {
    type: 'user',
    session_id: 'session-1',
    message: { role: 'user', content: 'Hello' },
  },
})

await client.stop()

Example: Connect to Local Server

const client = new ClaudeAgentClient({
  connectionUrl: 'http://localhost:3000',
  anthropicApiKey: process.env.ANTHROPIC_API_KEY,
})

await client.start()

For more details, see packages/client/README.md.

Server API Reference

Note: If you're using the @dzhng/claude-agent library, you don't need to interact with these endpoints directly. The client handles configuration and WebSocket connections for you. This section is for advanced users who want to connect to the server directly or build their own client.

The server runs on http://localhost:3000 (or your E2B sandbox URL) with:

• Config endpoint: http://localhost:3000/config
• WebSocket endpoint: ws://localhost:3000/ws

Configuration API

POST /config

Set the configuration for the Claude Agent SDK query:

type QueryConfig = {
  agents?: Record<string, AgentDefinition>
  allowedTools?: string[]
  systemPrompt?:
    | string
    | {
        type: 'preset'
        preset: 'claude_code'
        append?: string
      }
  model?: string
  anthropicApiKey?: string
}

Example:

curl -X POST http://localhost:3000/config \
  -H "Content-Type: application/json" \
  -d '{
    "systemPrompt": "You are a helpful assistant.",
    "allowedTools": ["read_file", "write_file"],
    "anthropicApiKey": "sk-ant-...",
    "model": "claude-3-5-sonnet-20241022",
    "agents": {
      "myAgent": {
        "name": "My Custom Agent",
        "description": "A custom agent"
      }
    }
  }'

Response:

{
  "success": true,
  "config": {
    "systemPrompt": "You are a helpful assistant.",
    "allowedTools": ["read_file", "write_file"],
    "agents": { ... }
  }
}

GET /config

Get the current configuration:

curl http://localhost:3000/config

Response:

{
  "config": {
    "systemPrompt": "You are a helpful assistant.",
    "allowedTools": ["read_file", "write_file"]
  }
}

WebSocket API

Connecting

Connect to the WebSocket endpoint:

const ws = new WebSocket('ws://localhost:3000/ws')

Note: The server only accepts one active connection at a time. If another client is already connected, new connection attempts will be rejected with an error message.

Message Format

Sending Messages (Client → Server)

type WSInputMessage =
  | {
      type: 'user_message'
      data: SDKUserMessage
    }
  | {
      type: 'interrupt'
    }

User Message:

Send a wrapped SDKUserMessage:

{
  "type": "user_message",
  "data": {
    "type": "user",
    "session_id": "your-session-id",
    "parent_tool_use_id": null,
    "message": {
      "role": "user",
      "content": "Hello, Claude!"
    }
  }
}

Structure:

• type: Must be "user_message"
• data: An SDKUserMessage object containing:
  • type: Must be "user"
  • session_id: Your session identifier (string)
  • message: An object with role and content
    • role: Must be "user"
    • content: The message content (string)
  • parent_tool_use_id: Optional, for tool use responses
  • uuid: Optional, message UUID (auto-generated if not provided)

Interrupt Message:

Send an interrupt to stop the current agent operation:

{
  "type": "interrupt"
}

Receiving Messages (Server → Client)

type WSOutputMessage =
  | { type: 'connected' }
  | { type: 'sdk_message'; data: unknown }
  | { type: 'error'; error: string }

Connection confirmation:

{
  "type": "connected"
}

SDK messages (responses from Claude):

{
  "type": "sdk_message",
  "data": {
    "type": "assistant",
    "session_id": "...",
    "message": {
      /* Claude's response */
    }
  }
}

Error messages:

{
  "type": "error",
  "error": "Error description"
}

Architecture

The server is a simple 1-to-1 relay between a single WebSocket client and the Claude Agent S