Amp SDK

The Amp SDK allows you to programmatically use the Amp agent in your TypeScript or Python programs.

Why use the Amp SDK?

The Amp SDK offers the following functionality:

  • Stream Inputs: Send messages one-by-one to the Amp agent
  • Stream Outputs: Receive structured JSON responses (system, assistant, result) while the agent runs
  • Multi-turn Conversations: Maintain back-and-forth interactions across multiple inference calls
  • Thread Continuity: Continue an existing thread (latest or by ID) to build stateful agent workflows
  • Programmatic Settings: Configure working directories, settings, and tools without user prompts — ideal for automation
  • MCP Integration: specify which MCP servers are available for a given session
  • Custom Skills: Define and use custom agent skills to extend Amp’s functionality

What can you build?

Here are some examples of what you can build with the Amp SDK:

Development Tools

  • Code Review Agent: Automated pull request analysis and feedback
  • Documentation Generator: Create and maintain project documentation
  • Test Automation: Generate and execute test suites
  • Migration Assistant: Help upgrade codebases and refactor legacy code

Workflow Automation

  • CI/CD Integration: Smart build and deployment pipelines
  • Issue Triage: Automatically categorize and prioritize bug reports
  • Code Quality Monitoring: Continuous analysis of code health metrics
  • Release Management: Automated changelog generation and version bumping

Quick Start

Installation

# Install the Amp SDK using npm
npm install @ampcode/sdk

# or yarn
yarn add @ampcode/sdk

# Install or upgrade the Amp CLI used by the SDK (optional if a compatible version is already installed)
npx -y @ampcode/sdk install

For SDK users who need to use Amp before Amp Neo, install the legacy release @ampcode/sdk@0.1.0-20260528044221-ge0e19fa:

npm install @ampcode/sdk@0.1.0-20260528044221-ge0e19fa

For TypeScript SDK users, Amp CLI must be at least the version pinned by the installed SDK release. If your organization already installs Amp CLI through Homebrew, Artifactory, or another internal channel, you can keep using it as long as its version is new enough.

Once installed, add your access token to the environment. You can access your access token in Security Settings.

export AMP_API_KEY=sgamp_your_access_token_here

If you already have the Amp CLI installed locally, you can log in using the following command amp login.

Run Your First Amp Command

Now that you have the SDK installed and your access token set up, you can start using Amp with the execute() function:

import { execute } from '@ampcode/sdk'

// Simple execution - get the final result
for await (const message of execute({ prompt: 'What files are in this directory?' })) {
	if (message.type === 'result' && !message.is_error) {
		console.log('Result:', message.result)
		break
	}
}

The execute() function only requires that you provide a prompt to get started. The SDK streams messages as the agent works, letting you handle responses and integrate them directly into your application.

Core Concepts

Message Streaming

The SDK streams different types of messages as your agent executes:

for await (const message of execute({ prompt: 'Run tests' })) {
	if (message.type === 'system') {
		// Session info, available tools, MCP servers
		console.log('Available tools:', message.tools)
	} else if (message.type === 'assistant') {
		// AI responses and tool usage
		console.log('Assistant is working...')
	} else if (message.type === 'result') {
		console.log(message.is_error ? `Failed: ${message.error}` : `Done: ${message.result}`)
	}
}

Simple Result Extraction

When you just need the final result without handling streaming:

async function getResult(prompt: string): Promise<string> {
	for await (const message of execute({ prompt, options: {} })) {
		if (message.type === 'result') {
			if (message.is_error) {
				throw new Error(message.error)
			}
			return message.result
		}
	}
	throw new Error('No result received')
}

// Usage
try {
	const result = await getResult('List all TypeScript files in this project')
	console.log('Found files:', result)
} catch (error) {
	console.error('Failed:', error instanceof Error ? error.message : String(error))
}

Thread Continuity

Continue conversations across multiple interactions:

// Continue the most recent conversation
for await (const message of execute({
	prompt: 'What was the last error you found?',
	options: { continue: true },
})) {
	if (message.type === 'result' && !message.is_error) {
		console.log(message.result)
	}
}

// Continue a specific thread by ID
for await (const message of execute({
	prompt: 'Can you update that code we discussed?',
	options: { continue: 'T-abc123-def456' },
})) {
	if (message.type === 'result' && !message.is_error) {
		console.log(message.result)
	}
}

Common Configuration

Working Directory

Specify where Amp should run:

for await (const message of execute({
	prompt: 'Refactor the auth module',
	options: { cwd: './my-project' },
})) {
	// Process messages...
}

Enable Debug Logging

See what’s happening under the hood:

for await (const message of execute({
	prompt: 'Analyze this project',
	options: {
		logLevel: 'debug', // Shows CLI command in console
		logFile: './amp-debug.log', // Optional: write logs to file
	},
})) {
	// Process messages
}

Agent Mode

Select which agent mode to use. The mode controls the model, system prompt, and tool selection:

for await (const message of execute({
	prompt: 'Quickly fix this typo',
	options: {
		mode: 'low', // Use low mode for faster responses
	},
})) {
	// Process messages
}

Available modes:

  • low: Fast, low-cost mode for small, well-defined tasks
  • medium (default): Balances quality, speed, and cost for most tasks
  • high: Deep reasoning for difficult tasks that can tolerate more time and cost
  • ultra: For the hardest open-ended tasks, when maximum capability matters more than speed or cost

Reasoning Effort

Set model reasoning effort for supported modes:

for await (const message of execute({
	prompt: 'Think carefully and explain your plan before coding.',
	options: {
		mode: 'medium',
		effort: 'high',
	},
})) {
	if (message.type === 'result' && !message.is_error) {
		console.log(message.result)
		break
	}
}

Available effort levels:

  • none
  • minimal
  • low
  • medium
  • high
  • xhigh
  • max

Thread Labels

Add labels to threads created by execute():

for await (const message of execute({
	prompt: 'Summarize this repo',
	options: {
		labels: ['sdk', 'summary'],
	},
})) {
	if (message.type === 'result' && !message.is_error) {
		console.log(message.result)
		break
	}
}

Thread Titles

Set the title when you create a thread. You cannot use title when continuing a thread.

for await (const message of execute({
	prompt: 'Review pull request 42',
	options: {
		executor: 'orb',
		title: 'Review acme/widgets#42',
	},
})) {
	if (message.type === 'result' && !message.is_error) {
		console.log(message.result)
		break
	}
}

Thread Visibility

Control who can see threads created by execute():

for await (const message of execute({
	prompt: 'Analyze this private codebase',
	options: {
		visibility: 'private', // Visible only to you (and workspace admins if you're in a workspace)
	},
})) {
	if (message.type === 'result' && !message.is_error) {
		console.log(message.result)
		break
	}
}

Available visibility levels:

  • workspace (default): Visible to all workspace members
  • private: Visible only to you (and workspace admins if you’re in a workspace)
  • unlisted: Visible to anyone with the link
  • group: Visible to members of your user group (Enterprise)

Tool Permissions

Use a custom plugin to control tool use. Workspace admins can distribute it as a global workspace plugin.

Advanced Usage

Interactive Progress Tracking

For building user interfaces that show real-time progress:

async function executeWithProgress(prompt: string) {
	console.log('Starting task...')

	for await (const message of execute({ prompt })) {
		if (message.type === 'system' && message.subtype === 'init') {
			console.log('Tools available:', message.tools.join(', '))
		} else if (message.type === 'assistant') {
			// Show tool usage or assistant responses
			for (const content of message.message.content) {
				if (content.type === 'tool_use') {
					console.log(`Using ${content.name}...`)
				} else if (content.type === 'text') {
					console.log('Assistant:', content.text.slice(0, 100) + '...')
				}
			}
		} else if (message.type === 'result') {
			if (message.is_error) {
				console.log('Failed:', message.error)
			} else {
				console.log('Completed successfully!')
				console.log(message.result)
			}
		}
	}
}

Cancellation and Timeouts

Handle long-running operations gracefully:

async function executeWithTimeout(prompt: string, timeoutMs = 30000) {
	const signal = AbortSignal.timeout(timeoutMs)

	try {
		for await (const message of execute({
			prompt,
			signal,
			options: {},
		})) {
			if (message.type === 'result') {
				if (message.is_error) {
					throw new Error(message.error)
				}
				return message.result
			}
		}
	} catch (error) {
		if (signal.aborted) {
			throw new Error(`Operation timed out after ${timeoutMs}ms`)
		}
		throw error
	}
}

MCP Integration

Extend Amp’s capabilities with custom tools and data sources:

import { execute, type MCPConfig } from '@ampcode/sdk'

const mcpConfig: MCPConfig = {
	playwright: {
		command: 'npx',
		args: ['-y', '@playwright/mcp@latest', '--headless'],
		env: { NODE_ENV: 'production' },
	},
	database: {
		command: 'node',
		args: ['./custom-mcp-server.js'],
		// ${VAR} placeholders are expanded from Amp's environment at server startup
		env: { DB_CONNECTION_STRING: '${DATABASE_URL}' },
	},
}

for await (const message of execute({
	prompt: 'Test the login flow on staging environment',
	options: { mcpConfig },
})) {
	if (message.type === 'system') {
		console.log(
			'MCP Servers:',
			message.mcp_servers.map((s) => `${s.name}: ${s.status}`),
		)
	}
	// Handle other messages...
}

To learn more about extending Amp with MCP servers, see MCP.

Multi-turn Conversations

Build streaming conversations using async generators:

import { execute, createUserMessage } from '@ampcode/sdk'

async function* generateMessages() {
	yield createUserMessage('Start analyzing the codebase')

	// Wait for some condition or user input
	await new Promise((resolve) => setTimeout(resolve, 1000))

	yield createUserMessage('Now focus on the authentication module')
}

for await (const message of execute({
	prompt: generateMessages(),
})) {
	if (message.type === 'result' && !message.is_error) {
		console.log(message.result)
	}
}

Settings File Configuration

Configure Amp’s behavior with a settings file, like the settings.json. You can provide Amp with a custom settings file you have saved in your project:

import { execute } from '@ampcode/sdk'

// Use a custom settings file
for await (const message of execute({
	prompt: 'Deploy the application',
	options: {
		settingsFile: './settings.json',
		logLevel: 'debug',
	},
})) {
	// Handle messages...
}

Example settings.json:

{
	"amp.mcpServers": {
		"playwright": {
			"command": "npx",
			"args": ["-y", "@playwright/mcp@latest", "--headless", "--isolated"]
		}
	},
	"amp.tools.disable": ["web_search", "mcp__playwright__browser_resize"]
}

To find all available settings, see Configuration.

Custom Skills

Load custom skills from a specified directory:

for await (const message of execute({
	prompt: 'Use my custom deployment skill',
	options: {
		skills: './my-skills', // Path to custom skills directory
	},
})) {
	// Process messages
}

To learn more about creating custom skills, see Skills.