ai-provider-anthropic-sdk

Anthropic SDK Patterns

Quick Guide: Use the official @anthropic-ai/sdk package to interact with Claude models directly. Use client.messages.create() for single-turn and multi-turn conversations. Use client.messages.stream() for streaming with event-based consumption. max_tokens is always required. Content blocks are typed unions (text, tool_use, thinking). Use client.messages.parse() with zodOutputFormat() for structured outputs. Tool use requires a tool-result loop -- Claude returns tool_use blocks, you execute the tool and send back tool_result blocks. Extended thinking adds thinking content blocks before the response.

---

<critical_requirements>

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST always provide max_tokens in every messages.create() / messages.stream() call -- it is required and has no default)

(You MUST handle the stop_reason field to detect end_turn, max_tokens, tool_use, and stop_sequence -- ignoring it causes silent truncation or broken tool loops)

(You MUST iterate over response.content blocks (not assume a single text block) -- responses can contain text, tool_use, and thinking blocks mixed together)

(You MUST handle errors using Anthropic.APIError and its subclasses -- never use bare catch blocks without error type checking)

(You MUST never hardcode API keys -- always use environment variables via process.env.ANTHROPIC_API_KEY)

</critical_requirements>

---

Auto-detection: Anthropic, @anthropic-ai/sdk, client.messages.create, client.messages.stream, client.messages.parse, client.messages.countTokens, client.messages.batches, ANTHROPIC_API_KEY, claude-sonnet, claude-opus, claude-haiku, ContentBlock, ToolUseBlock, tool_use, tool_result, thinking, budget_tokens, adaptive, cache_control, zodOutputFormat, betaZodTool, toolRunner

When to use:

• Building applications that call Claude models directly (Opus, Sonnet, Haiku families)
• Implementing streaming chat responses with event-based text accumulation
• Using tool use / function calling where Claude decides which tools to invoke
• Processing images, PDFs, or documents alongside text prompts
• Enabling extended thinking for complex reasoning tasks
• Extracting structured data from responses with Zod schema validation
• Caching large system prompts or conversation prefixes for cost savings
• Running batch jobs for high-volume, asynchronous processing
• Counting tokens before sending requests for cost estimation

Key patterns covered:

• Client initialization and configuration (retries, timeouts, API key)
• Messages API (messages.create, system prompts, multi-turn conversations)
• Streaming with .stream() helper and stream: true low-level SSE
• Tool use / function calling (tools array, tool_use / tool_result content blocks)
• Vision (base64 images, URL images, PDFs/documents)
• Extended thinking (thinking config, budget_tokens, thinking content blocks)
• Structured outputs (zodOutputFormat, messages.parse, output_config)
• Prompt caching (cache_control: { type: "ephemeral" })
• Batch API (messages.batches.create)
• Token counting (messages.countTokens)
• Error handling, retries, and production best practices

When NOT to use:

• Multi-provider applications where you need to switch between multiple LLM providers -- use a unified provider SDK instead
• React-specific chat UI hooks (useChat, useCompletion) -- use a framework-integrated AI SDK
• When you need a higher-level agent framework -- consider the Claude Agent SDK (@anthropic-ai/claude-agent-sdk)

---

Examples Index

• Core: Setup & Configuration -- Client init, production config, error handling, token counting
• Streaming -- .stream() helper, stream: true SSE, event types, abort
• Tool Use / Function Calling -- Tool definitions, tool loops, parallel tool calls, automated tool runner
• Vision & Documents -- Base64 images, URL images, PDFs, multi-modal
• Extended Thinking -- Thinking config, streaming thinking, thinking with tool use
• Quick API Reference -- Model IDs, method signatures, error types, streaming events, content block types

---

Philosophy

The official Anthropic SDK provides direct, typed access to the Claude API. It is auto-generated from Anthropic's API specification using Stainless, giving you the exact API surface that Anthropic documents with full TypeScript types.

Core principles:

1. Content blocks, not strings -- Responses are arrays of typed content blocks (TextBlock, ToolUseBlock, ThinkingBlock), not plain strings. Always iterate over response.content and switch on block.type.
2. Explicit resource limits -- max_tokens is always required. There is no default. The API will reject requests without it.
3. Tool use is a conversation loop -- When stop_reason === "tool_use", Claude is requesting you execute a tool. You must send the result back as a tool_result content block to continue the conversation.
4. Built-in resilience -- The SDK retries 2 times by default on 429, 409, 408, 529, and 5xx errors with exponential backoff.
5. Streaming as a first-class pattern -- Use .stream() for an event-based API with .on("text", ...), or stream: true for raw SSE iteration.

When to use the Anthropic SDK directly:

• You only use Claude models and want the simplest, most direct integration
• You need access to Anthropic-specific features (extended thinking, prompt caching, batch API)
• You want minimal dependencies and zero abstraction overhead
• You need the latest API features on day one

When NOT to use:

• You need to switch between multiple LLM providers -- use a unified provider SDK
• You want React-specific chat UI hooks -- use a framework-integrated AI SDK
• You want a higher-level agent framework -- consider the Claude Agent SDK

---

Core Patterns

Pattern 1: Client Setup

Initialize the Anthropic client. It auto-reads ANTHROPIC_API_KEY from the environment.

// lib/anthropic.ts -- basic setup
import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();
export { client };

// lib/anthropic.ts -- production configuration
const TIMEOUT_MS = 30_000;
const MAX_RETRIES = 3;
const client = new Anthropic({ timeout: TIMEOUT_MS, maxRetries: MAX_RETRIES });

Why good: Minimal setup, env var auto-detected, named constants for production settings

// BAD: Hardcoded API key
const client = new Anthropic({ apiKey: "sk-ant-api03-..." });

Why bad: Hardcoded keys get committed to version control, causing security breaches

See: examples/core.md for per-request overrides, error handling patterns, token counting

---

Pattern 2: Messages API

All interactions use client.messages.create(). max_tokens is always required.

const MAX_TOKENS = 1024;

const message = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: MAX_TOKENS,
  system: "You are a helpful coding assistant.",
  messages: [{ role: "user", content: "Explain TypeScript generics." }],
});

// Response is an array of content blocks -- iterate, don't assume
for (const block of message.content) {
  if (block.type === "text") {
    console.log(block.text);
  }
}

Why good: Named constant for max_tokens, system prompt separated from messages, content blocks iterated

// BAD: Assuming content is a single text string
const text = message.content[0].text; // Crashes if block is tool_use or thinking

Why bad: Content can contain multiple blocks of different types -- direct index access without type checking crashes at runtime

See: examples/core.md for multi-turn conversations, system prompts, token tracking

---

Pattern 3: Streaming

Use .stream() for event-based streaming with text accumulation helpers.

const MAX_TOKENS = 1024;

const stream = client.messages.stream({
  model: "claude-sonnet-4-6",
  max_tokens: MAX_TOKENS,
  messages: [{ role: "user", content: "Explain async/await." }],
});

stream.on("text", (text) => {
  process.stdout.write(text);
});

const finalMessage = await stream.finalMessage();

Why good: Event-based API handles accumulation, finalMessage() gives the complete response object

// BAD: Using stream: true without consuming events
const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: MAX_TOKENS,
  messages: [{ role: "user", content: "Hello" }],
  stream: true,
});
// Response is an async iterable, not a Message -- must iterate

Why bad: stream: true returns an async iterable of raw SSE events, not a Message. Treating it as a Message silently breaks.

See: examples/streaming.md for raw SSE iteration, abort, stream events, streaming with thinking

---

Pattern 4: Tool Use / Function Calling

Define tools Claude can invoke. Handle the tool_use -> tool_result conversation loop.

const tools: Anthropic.Messages.Tool[] = [
  {
    name: "get_weather",
    description: "Get current weather for a location",
    input_schema: {
      type: "object" as const,
      properties: {
        location: { type: "string", description: "City name" },
      },
      required: ["location"],
    },
  },
];

const MAX_TOKENS = 1024;

const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: MAX_TOKENS,
  tools,
  messages: [{ role: "user", content: "Weather in Paris?" }],
});

// Check stop_reason to know if Claude wants to call a tool
if (response.stop_reason === "tool_use") {
  const toolBlock = response.content.find(
    (block): block is Anthropic.Messages.ToolUseBlock =>
      block.type === "tool_use",
  );
  if (toolBlock) {
    console.log(`Call ${toolBlock.name} with:`, toolBlock.input);
  }
}

Why good: Typed tool definitions, stop_reason checked, type guard for ToolUseBlock

// BAD: Not checking stop_reason, not sending tool_result back
const response = await client.messages.create({
  /* ... with tools */
});
console.log(response.content[0]); // May be a tool_use block, not text!

Why bad: When Claude wants to call a tool, there is no text content -- only tool_use blocks. You must execute the tool and send back a tool_result to get the final answer.

See: examples/tool-use.md for complete tool loops, parallel tool calls, automated tool runner

---

Pattern 5: Vision & Documents

Pass images and PDFs as content blocks alongside text.

import { readFileSync } from "node:fs";

const MAX_TOKENS = 1024;
const imageData = readFileSync("photo.jpg").toString("base64");

const message = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: MAX_TOKENS,
  messages: [
    {
      role: "user",
      content: [
        {
          type: "image",
          source: { type: "base64", media_type: "image/jpeg", data: imageData },
        },
        { type: "text", text: "What's in this image?" },
      ],
    },
  ],
});

Why good: Multi-part content array, explicit media type, text and image combined in one message

See: examples/vision-documents.md for URL images, PDFs, multiple images

---

Pattern 6: Extended Thinking

Enable extended thinking for complex reasoning. Responses include thinking content blocks. Use adaptive thinking on Opus 4.6 and Sonnet 4.6 (recommended). Use manual budget_tokens on older models.

const MAX_TOKENS = 16_000;

// Adaptive thinking (recommended for 4.6 models)
const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: MAX_TOKENS,
  thinking: { type: "adaptive" },
  messages: [
    { role: "user", content: "Prove there are infinitely many primes." },
  ],
} as unknown as Anthropic.MessageCreateParamsNonStreaming);

for (const block of response.content) {
  if (block.type === "thinking") {
    console.log("Thinking:", block.thinking);
  } else if (block.type === "text") {
    console.log("Answer:", block.text);
  }
}

Why good: Adaptive thinking lets Claude decide how much to reason, iterates content blocks, handles both thinking and text blocks

// Manual thinking (deprecated on 4.6 models, required on older models)
const THINKING_BUDGET = 10_000;

const response = await client.messages.create({
  model: "claude-sonnet-4-5",
  max_tokens: MAX_TOKENS,
  thinking: { type: "enabled", budget_tokens: THINKING_BUDGET },
  messages: [
    { role: "user", content: "Prove there are infinitely many primes." },
  ],
});

Note: The TypeScript SDK does not yet have "adaptive" in its type definitions. The as unknown as Anthropic.MessageCreateParamsNonStreaming assertion is required until the SDK types are updated.

See: examples/extended-thinking.md for streaming thinking, thinking with tools, display options

---

Pattern 7: Structured Outputs

Use zodOutputFormat() and messages.parse() for type-safe structured responses.

import { zodOutputFormat } from "@anthropic-ai/sdk/helpers/zod";
import { z } from "zod";

const ContactInfo = z.object({
  name: z.string(),
  email: z.string(),
  topics: z.array(z.string()),
});

const MAX_TOKENS = 1024;

const response = await client.messages.parse({
  model: "claude-sonnet-4-6",
  max_tokens: MAX_TOKENS,
  messages: [
    {
      role: "user",
      content:
        "Extract info: John ([email protected]) asked about billing and API limits.",
    },
  ],
  output_config: { format: zodOutputFormat(ContactInfo) },
});

const parsed = response.parsed_output; // Fully typed: { name, email, topics }

Why good: Auto-converts Zod schema, validates output, fully typed result

See: examples/core.md for raw JSON schema, combined with tool use

---

Pattern 8: Prompt Caching

Cache large system prompts and conversation prefixes for cost savings.

const MAX_TOKENS = 1024;

const response = await client.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: MAX_TOKENS,
  system: [
    {
      type: "text",
      text: "You are a legal document analyst.",
    },
    {
      type: "text",
      text: largeDocumentText, // 50+ pages of legal text
      cache_control: { type: "ephemeral" },
    },
  ],
  messages: [{ role: "user", content: "What are the key terms?" }],
});

// Check cache performance
console.log("Cache read tokens:", response.usage.cache_read_input_tokens);
console.log("Cache write tokens:", response.usage.cache_creation_input_tokens);

Why good: Cache breakpoint on the large static content, cache metrics tracked

See: reference.md for cache pricing, TTL options, automatic caching

---

Pattern 9: Error Handling

Always catch Anthropic.APIError and its subclasses. Re-throw unexpected errors.

try {
  const message = await client.messages.create({
    model: "claude-sonnet-4-6",
    max_tokens: 1024,
    messages: [{ role: "user", content: "Hello" }],
  });
} catch (error) {
  if (error instanceof Anthropic.APIError) {
    console.error(`API Error [${error.status}]: ${error.message}`);

    if (error instanceof Anthropic.RateLimitError) {
      console.error("Rate limited -- SDK will auto-retry 2 times");
    }
    if (error instanceof Anthropic.AuthenticationError) {
      throw new Error("Invalid API key. Check ANTHROPIC_API_KEY.");
    }
  } else {
    throw error; // Re-throw non-API errors
  }
}

Why good: Specific error types, status code access, re-throws unexpected errors

See: examples/core.md for full error hierarchy, stream error handling

---

Performance Optimization

Model Selection for Cost/Speed

Most capable, complex reasoning -> claude-opus-4-6  (1M context, 128K output)
General purpose, best value      -> claude-sonnet-4-6 (1M context, 64K output)
Fast + cheap, simple tasks       -> claude-haiku-4-5  (200K context, 64K output)
Extended thinking                -> claude-sonnet-4-6 or claude-opus-4-6 (use adaptive thinking)
Vision / multimodal              -> claude-sonnet-4-6 or claude-opus-4-6
Batch processing                 -> Any model at 50% batch discount

Key Optimization Patterns

• Track token usage via message.usage for cost visibility (input_tokens, output_tokens)
• Check stop_reason === "max_tokens" to detect truncated output
• Use prompt caching for large system prompts -- cache reads cost 0.1x base input price
• Use messages.countTokens() before sending to estimate costs
• Use Batch API for high-volume async jobs at 50% cost reduction
• Use AbortController to cancel long-running requests
• Set temperature: 0 for deterministic output when caching matters

---

<decision_framework>

Decision Framework

Which Model to Choose

What is your task?
+-- Complex reasoning / analysis    -> claude-opus-4-6
+-- General purpose (best balance)  -> claude-sonnet-4-6
+-- Fast + cheap, high throughput   -> claude-haiku-4-5
+-- Extended thinking needed        -> claude-sonnet-4-6 (or opus-4-6 with adaptive thinking)
+-- Vision / image analysis         -> claude-sonnet-4-6 or claude-opus-4-6
+-- Batch processing                -> Any model (50% discount)

Streaming vs Non-Streaming

Is the response user-facing?
+-- YES -> Use streaming (client.messages.stream())
|   +-- Need event-level control? -> .on("text", ...) + .on("contentBlock", ...)
|   +-- Just want final message?  -> stream.finalMessage() (avoids HTTP timeouts on large responses)
+-- NO -> Use non-streaming (client.messages.create())
    +-- Background processing  -> messages.create()
    +-- Structured output      -> messages.parse()
    +-- High volume            -> Batch API

When to Use Extended Thinking

Does the task require multi-step reasoning?
+-- YES -> Which model?
|   +-- Opus 4.6 or Sonnet 4.6? -> Use adaptive: thinking: { type: "adaptive" }
|   |   +-- Control depth?       -> Add output_config: { effort: "high" | "medium" | "low" }
|   |   +-- Opus only max depth? -> effort: "max"
|   +-- Older models?            -> Manual: thinking: { type: "enabled", budget_tokens: N }
+-- NO -> Standard messages.create() is sufficient (omit thinking param or type: "disabled")

</decision_framework>

---

<red_flags>

RED FLAGS

High Priority Issues:

• Not providing max_tokens (request will be rejected -- it has no default)
• Hardcoding API keys instead of using environment variables (security breach risk)
• Treating response.content as a string instead of iterating content blocks (crashes on tool_use or thinking blocks)
• Not checking stop_reason for "tool_use" (breaks function calling flows -- Claude is waiting for tool results)
• Using bare catch blocks without checking Anthropic.APIError (hides API-specific error information)

Medium Priority Issues:

• Not setting maxRetries / timeout for production deployments (default timeout is 10 minutes, which may be too long)
• Ignoring stop_reason === "max_tokens" (response was truncated but you are using it as complete)
• Ignoring usage data (no cost visibility or budget tracking)
• Not sending thinking blocks back in multi-turn conversations when using extended thinking (Claude loses reasoning context)
• Changing thinking parameters between turns in a tool use loop (invalidates message cache, causes errors)

Common Mistakes:

• Using system as a message role instead of the top-level system parameter (there is no system role in messages -- use the system parameter)
• Assuming response.content has exactly one block (it can have multiple text, tool_use, and thinking blocks)
• Not passing tool_result back after a tool_use response (Claude cannot continue without it)
• Using max_completion_tokens instead of max_tokens (the Anthropic API uses max_tokens, not max_completion_tokens)
• Using response_format instead of output_config for structured outputs (wrong parameter name)
• Forgetting that budget_tokens must be less than max_tokens (except with interleaved thinking)

Gotchas & Edge Cases:

• The SDK auto-retries on 429 (rate limit), 529 (overloaded), 408 (timeout), 409 (conflict), and 5xx errors -- 2 retries by default with exponential backoff. Disable with maxRetries: 0.
• client.messages.stream() returns a MessageStream with event helpers. client.messages.create({ stream: true }) returns a raw async iterable of SSE events. They are different APIs.
• When using extended thinking with tool use, you must include the thinking blocks unmodified when sending conversation history back. Omitting or modifying them causes errors.
• tool_choice: { type: "any" } forces Claude to call a tool but cannot be used with extended thinking. Only "auto" and "none" work with thinking enabled.
• Prompt caching requires a minimum of 1024-4096 tokens (model-dependent) to be cacheable. Small prompts will not be cached.
• Cache breakpoints on messages are invalidated when thinking parameters change between requests. System prompt cache is preserved.
• budget_tokens is deprecated on both Claude Opus 4.6 and Sonnet 4.6 -- use thinking: { type: "adaptive" } instead. budget_tokens still works but will be removed in a future release.
• The display field on thinking config controls whether thinking text is returned: "summarized" (default) or "omitted" (only signature, faster streaming).
• Adaptive thinking automatically enables interleaved thinking (thinking between tool calls). Manual mode on Sonnet 4.6 requires the interleaved-thinking-2025-05-14 beta header for interleaved thinking.
• The effort parameter (output_config: { effort: "high" | "medium" | "low" | "max" }) works with adaptive thinking to control thinking depth. "max" is Opus 4.6 only.
• The TypeScript SDK does not yet include "adaptive" in its type definitions -- use a type assertion when passing thinking: { type: "adaptive" }.
• Multi-turn conversations require you to include the full assistant response (all content blocks) in the conversation history, not just the text.
• Batch API requests have a 24-hour completion window. Use messages.batches.results() to retrieve completed results.

</red_flags>

---

<critical_reminders>

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST always provide max_tokens in every messages.create() / messages.stream() call -- it is required and has no default)

(You MUST handle the stop_reason field to detect end_turn, max_tokens, tool_use, and stop_sequence -- ignoring it causes silent truncation or broken tool loops)

(You MUST iterate over response.content blocks (not assume a single text block) -- responses can contain text, tool_use, and thinking blocks mixed together)

(You MUST handle errors using Anthropic.APIError and its subclasses -- never use bare catch blocks without error type checking)

(You MUST never hardcode API keys -- always use environment variables via process.env.ANTHROPIC_API_KEY)

Failure to follow these rules will produce broken tool loops, silent truncation, security vulnerabilities, or untyped AI integrations.

</critical_reminders>