| name | TypeScript SDK Reference |
| description | Comprehensive reference for the TypeScript Agent SDK with functions, types, and interfaces for programmatic Claude Code interactions |
| allowed-tools | Read, Write, Edit, Task |
TypeScript SDK Reference
A comprehensive reference for building applications with the Anthropic Agent SDK (TypeScript), enabling programmatic interactions with Claude Code and custom tool integrations.
Installation
npm install @anthropic-ai/claude-agent-sdk
Quick Start
Core Functions
| Function | Purpose | Key Use Case |
|---|---|---|
query() |
Primary interface for Claude Code interactions | Send prompts, stream responses, manage multi-turn conversations |
tool() |
Define MCP tools with type-safe schemas | Create custom tools for agents using Zod validation |
createSdkMcpServer() |
Create in-process MCP servers | Host tools natively in your application |
Example: Basic Query
import { query } from '@anthropic-ai/claude-agent-sdk';
const result = query({
prompt: "Analyze this code and suggest improvements",
options: {
allowedTools: ['Read', 'Grep'],
model: 'claude-opus'
}
});
for await (const message of result) {
if (message.type === 'assistant') {
console.log(message.message.content);
}
}
Main Functions Reference
query()
Streams Claude Code responses with full message type support. Returns an async generator of SDKMessage objects.
Parameters:
prompt: string or async iterable of messages (streaming mode)options: Configuration object (see Options below)
Returns: Query object (extends AsyncGenerator<SDKMessage, void>)
Key Methods:
.interrupt()– Interrupt streaming (streaming mode only).setPermissionMode(mode)– Change permissions dynamically (streaming mode only)
tool()
Creates type-safe MCP tool definitions using Zod schemas. Handlers receive validated inputs and must return CallToolResult.
Parameters:
name: Tool identifierdescription: Natural language descriptioninputSchema: Zod schema defining inputshandler: Async function(args: z.infer<Schema>, extra?: unknown) => Promise<CallToolResult>
Returns: SdkMcpToolDefinition<Schema>
createSdkMcpServer()
Creates an in-process MCP server for hosting custom tools. Runs in the same Node.js process—no subprocess overhead.
Parameters:
name: Server identifierversion: Optional semantic versiontools: Array of tool definitions fromtool()
Returns: McpSdkServerConfigWithInstance (ready for mcpServers option)
Key Configuration Options
Essential Options
| Option | Type | Default | Purpose |
|---|---|---|---|
allowedTools |
string[] |
All available | Restrict which tools Claude can use |
disallowedTools |
string[] |
[] |
Explicitly block tools |
model |
string |
CLI default | Override Claude model (e.g., 'claude-opus') |
cwd |
string |
process.cwd() |
Working directory for operations |
abortController |
AbortController |
New instance | Control task cancellation |
Advanced Options
| Option | Type | Default | Purpose |
|---|---|---|---|
systemPrompt |
string or preset object | None | Custom system instructions or Claude Code preset |
agents |
Record<string, AgentDefinition> |
undefined |
Programmatically define subagents with custom prompts |
mcpServers |
Record<string, McpServerConfig> |
{} |
Connect stdio, SSE, HTTP, or SDK MCP servers |
settingSources |
('user' | 'project' | 'local')[] |
[] |
Load filesystem settings (CLAUDE.md, settings.json) |
permissionMode |
PermissionMode |
'default' |
'default', 'acceptEdits', 'bypassPermissions', or 'plan' |
maxThinkingTokens |
number |
undefined |
Token limit for extended thinking |
maxTurns |
number |
undefined |
Maximum conversation turns before stopping |
resume |
string |
undefined |
Resume a previous session by ID |
forkSession |
boolean |
false |
Fork to new session on resume instead of continuing |
Message Types Overview
The query() generator yields SDKMessage objects. Common types:
| Message Type | When Emitted | Key Fields |
|---|---|---|
'system' |
Session initialization | tools, model, permissionMode, slash_commands |
'user' |
User input sent | message (APIUserMessage), uuid |
'assistant' |
Claude responds | message (APIAssistantMessage with content/tool_use) |
'result' |
Query completes | subtype ('success' or error), usage, total_cost_usd |
'stream_event' |
Partial streaming data | event (only if includePartialMessages: true) |
See Message Types Reference for complete definitions.
Types Reference Overview
Core Configuration Types
Options– Query configuration with 20+ properties (Tools, MCP, permissions, execution)PermissionMode– Control execution permissions:'default','acceptEdits','bypassPermissions','plan'SettingSource– Filesystem config scope:'user','project','local'
Agent & Tool Types
AgentDefinition– Subagent config:description,prompt,tools,modeloverrideSdkMcpToolDefinition– Output oftool()function with Zod schema bindingMcpServerConfig– Union of stdio, SSE, HTTP, or SDK server configs
MCP Server Configuration
| Config Type | Transport | Use Case |
|---|---|---|
McpStdioServerConfig |
stdio (subprocess) | External tools, sandboxing |
McpSSEServerConfig |
Server-Sent Events | Remote tools, pub/sub patterns |
McpHttpServerConfig |
HTTP | REST-based tools, cloud integrations |
McpSdkServerConfigWithInstance |
In-process | Custom tools, no subprocess overhead |
Hook Types
HookEvent– Event names:PreToolUse,PostToolUse,Notification,UserPromptSubmit,SessionStart,SessionEnd, etc.HookCallback– Async function receiving hook input and returning decisionsHookInput– Union of all hook input types; each extendsBaseHookInput
See Types Reference for detailed type definitions.
Tool Input/Output Reference
Built-in Tool Inputs
All tool names map to input schemas:
| Tool | Purpose | Key Input |
|---|---|---|
Task |
Delegate to subagent | description, prompt, subagent_type |
Bash |
Run shell commands | command, optional timeout, run_in_background |
Read |
Read files/images/PDFs | file_path, optional offset/limit |
Write / Edit |
Modify files | file_path, content or old/new strings |
Glob |
Pattern matching | pattern, optional path |
Grep |
Regex search | pattern, optional filters (glob, type, -B/-A/-C) |
WebFetch |
Fetch & analyze URLs | url, prompt for AI analysis |
WebSearch |
Web search | query, optional domain filters |
NotebookEdit |
Jupyter notebooks | notebook_path, new_source, edit_mode |
TodoWrite |
Task tracking | todos array with status, content, activeForm |
See Tool Reference for complete input/output schemas and examples.
Common Patterns
Streaming with Progress
const result = query({
prompt: "Your task here",
options: {
includePartialMessages: true,
permissionMode: 'bypassPermissions'
}
});
for await (const msg of result) {
if (msg.type === 'stream_event') {
// Handle streaming events
console.log('Streaming:', msg.event);
} else if (msg.type === 'result') {
console.log('Cost:', msg.total_cost_usd, 'Usage:', msg.usage);
}
}
Custom MCP Tools
import { tool, createSdkMcpServer, query } from '@anthropic-ai/claude-agent-sdk';
import { z } from 'zod';
const myTool = tool(
'my-calculator',
'Add two numbers',
{ a: z.number(), b: z.number() },
async ({ a, b }) => ({
content: [{ type: 'text', text: `Result: ${a + b}` }]
})
);
const server = createSdkMcpServer({
name: 'my-tools',
tools: [myTool]
});
const result = query({
prompt: "Use my-calculator to add 5 + 3",
options: {
mcpServers: { 'my-tools': server }
}
});
Loading Project Settings
const result = query({
prompt: "Add a feature following project conventions",
options: {
systemPrompt: {
type: 'preset',
preset: 'claude_code' // Enables Claude Code system prompt
},
settingSources: ['project'], // Load .claude/settings.json & CLAUDE.md
allowedTools: ['Read', 'Write', 'Edit', 'Bash']
}
});
Related Resources
- Full API Reference – Complete type definitions and schemas
- Message Types Reference – SDKMessage union types and examples
- Types Reference – Detailed Options, AgentDefinition, PermissionMode docs
- Tool Reference – Input/output schemas for all built-in tools
- SDK Overview – Getting started guide
- Common Workflows – Step-by-step patterns
- CLI Reference – Command-line interface docs