feat: Support Anthropic as a model provider #374

chaynabors · 2025-12-29T22:08:56Z

Motivation

This PR integrates the Anthropic SDK to enable support for Claude models (e.g., Sonnet 3.5, Opus). While the SDK currently supports OpenAI and Bedrock, Anthropic is a highly requested provider, particularly for its strong reasoning capabilities, prompt caching, and native PDF support.

This implementation maps the SDK's abstract block types (including CachePointBlock and ReasoningBlock) to Anthropic's specific API requirements, allowing users to switch providers without rewriting message construction logic.

Resolves: #247

Public API Changes

A new AnthropicModel class is now available via a new entry point.

import { AnthropicModel } from '@strands-agents/sdk/anthropic'

// Basic instantiation
const model = new AnthropicModel({
  apiKey: process.env.ANTHROPIC_API_KEY, // Optional if set in env
  modelId: 'claude-sonnet-4-5-20250929',
  maxTokens: 4096,
})

// Support for Thinking (Reasoning)
const thinkingModel = new AnthropicModel({
  modelId: 'claude-sonnet-4-5-20250929',
  params: {
    thinking: { type: 'enabled', budget_tokens: 1024 }
  }
})

Prompt Caching behavior

The provider implements a lookahead strategy to support the SDK's CachePointBlock abstraction. It automatically attaches cache_control to the preceding content block as required by the Anthropic API.

// SDK Abstraction
const content = [
  new TextBlock(hugeContext),
  new CachePointBlock({ cacheType: 'ephemeral' }) 
]

// Transforms to Anthropic request:
// { type: 'text', text: hugeContext, cache_control: { type: 'ephemeral' } }

Use Cases

Cost Optimization: Leveraging CachePointBlock with Anthropic's prompt caching to reduce costs on long-context tasks (e.g., RAG or large system prompts).
Document Analysis: Native support for PDF processing via DocumentBlock.
Complex Reasoning: Access to Claude's "thinking" mode for tasks requiring extended chain-of-thought.

Implementation Notes

Binary Handling: Updated encodeBase64 utility to better handle Uint8Array inputs directly. This prevents stack overflow errors previously caused by spreading large arrays into String.fromCharCode when processing high-resolution images or PDFs.

src/models/__tests__/anthropic.test.ts

src/models/anthropic.ts

zastrowm · 2025-12-30T19:17:14Z

src/types/media.ts

+    // We convert bytes to binary string in chunks to satisfy btoa()
+    const CHUNK_SIZE = 0x8000 // 32k chunks
+    let binary = ''
+    for (let i = 0; i < input.length; i += CHUNK_SIZE) {


Could we use toBase64 for browser compatibility instead of chunking?

https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Uint8Array/toBase64

Since ⁨September 2025⁩, this feature works across the latest devices and browser versions. This feature might not work in older devices or browsers.

I don't know, you tell me.

zastrowm · 2025-12-30T19:17:47Z

test/integ/__fixtures__/model-providers.ts

  name: 'BedrockModel',
  get skip() {
-    return inject('provider-bedrock').shouldSkip
+    return inject('provider-bedrock')?.shouldSkip ?? true


Why switch to optional chaining here?

Inject can return undefined which crashes our test suite if it fails. Not sure why we weren't seeing a type error before this. You tell me if I misread it.

Not sure why we weren't seeing a type error before this.

How did the type error surface? If I pull locally and run:

npm run type-check

I'm not seeing an error with the original code

Inject can return undefined which crashes our test suite if it fails

In that case we might want it to fail - it would indicate that vite failed to inject values which means our tests are misconfigured.

Let's revert this otherwise

package.json

test/integ/anthropic.test.ts

strands-agent

Review Summary

👍 Great implementation! All tests pass and the code quality is excellent. There are 6 review threads that need addressing before this can merge. I've added specific guidance below to help you address each one.

📝 Action Items

1. Update PR Description ⭐ High Priority

Current: PR description is empty
Required: Follow the PR.md template

Suggested PR Description:

## Motivation

Anthropics's Claude models are widely used and offer competitive performance, especially for reasoning-heavy tasks. However, the SDK currently only supports Bedrock and OpenAI providers, requiring developers to either:
1. Use Claude through AWS Bedrock (adding infrastructure dependency)
2. Implement custom workarounds
3. Not use Strands SDK

Adding native Anthropic support enables direct Claude API integration without AWS dependency.

Resolves: #[issue-number if applicable]

## Public API Changes

Adds `AnthropicModel` class for direct Anthropic API integration:

```typescript
import { AnthropicModel } from '@strands-agents/sdk/anthropic'

const model = new AnthropicModel({
  modelId: 'claude-3-5-sonnet-20241022',
  apiKey: process.env.ANTHROPIC_API_KEY,
  temperature: 0.7,
  maxTokens: 4096
})

Key features:

Streaming and non-streaming responses
Tool/function calling support
Image input support (base64 encoding)
Context window overflow detection
Full TypeScript type safety

Follows OpenAI provider patterns for consistency.

Use Cases

Direct Claude API usage: Use Claude without AWS Bedrock infrastructure
Model comparison: Easily switch between OpenAI and Anthropic models
Cost optimization: Choose Claude models for specific use cases
Regional availability: Access Claude in regions where Bedrock isn't available


---

### 2. Make `@anthropic-ai/sdk` Optional Dependency ⭐ **High Priority**

**Current**: Listed in `dependencies`  
**Required**: Make optional (like `openai`)

**Why**: Users shouldn't need to install Anthropic SDK if they only use Bedrock/OpenAI.

**Changes needed**:

**a) `package.json`**: Move to `peerDependencies` + `peerDependenciesMeta`

```json
"peerDependencies": {
  "@anthropic-ai/sdk": "^0.38.0",
  "openai": "^4.73.0"
},
"peerDependenciesMeta": {
  "@anthropic-ai/sdk": {
    "optional": true
  },
  "openai": {
    "optional": true
  }
}

b) Add to package.json exports: (already have ./openai, add ./anthropic)

"./anthropic": {
  "types": "./dist/src/models/anthropic.d.ts",
  "default": "./dist/src/models/anthropic.js"
}

c) Remove from src/index.ts: Don't export from main entry point

d) Users import directly:

// Instead of:
import { AnthropicModel } from '@strands-agents/sdk'

// Users do:
import { AnthropicModel } from '@strands-agents/sdk/anthropic'

This matches the OpenAI pattern exactly! 👍

3. Fix Test File Name ⭐ High Priority

Current: test/integ/anthropic.ts
Required: test/integ/anthropic.test.ts

This ensures the test runner picks it up correctly.

4. API Key Callback Support 🔍 Needs Investigation

Question from @zastrowm: Does Anthropic SDK support API key callbacks like OpenAI?

OpenAI pattern (for reference):

import type { ApiKeySetter } from 'openai/client'

export interface OpenAIModelOptions {
  apiKey?: string | ApiKeySetter  // <-- Can be function
}

This enables:

Dynamic key rotation
Secret manager integration
Runtime credential refresh

Action needed: Check if @anthropic-ai/sdk supports similar functionality.

How to check:

Look at Anthropic SDK types: node_modules/@anthropic-ai/sdk/index.d.ts
Check their ClientOptions interface
If they support it, add the same pattern
If not, document in code comment: // Note: Anthropic SDK does not support dynamic API key callbacks yet

5. Use `toBase64` for Browser Compatibility 🔍 Investigate

Current: Custom chunking implementation in media.ts
Suggested: Use built-in Uint8Array.toBase64() method

Why: Better browser compatibility, less custom code.

Reference: MDN - Uint8Array.toBase64()

However: This is a newer API (2023), so check:

Browser compatibility requirements for the SDK
If targeting older browsers, current chunking may be necessary
Could add polyfill or feature detection

Suggested approach:

export function encodeBase64(data: Uint8Array): string {
  // Use native method if available (modern browsers/Node 22+)
  if (typeof data.toBase64 === 'function') {
    return data.toBase64()
  }
  
  // Fallback to chunking for older environments
  return chunkEncode(data)
}

6. Optional Chaining Change in `openai.ts` 🤔 Clarification Needed

Question from @zastrowm: Why was optional chaining added?

// What changed and why?
myProperty?.something

Please explain the reasoning in the PR thread. If it was unintentional, consider reverting to keep the PR focused on Anthropic support only.

✅ What's Already Great

🎉 All CI checks passing (10/10 green!)
📝 Comprehensive tests: Unit + integration tests
🎯 Follows patterns: Matches OpenAI provider structure
🔒 Type-safe: Full TypeScript support
⚡ Streaming support: Handles streaming responses correctly
🛠️ Tool calling: Implements function/tool use
🖼️ Image support: Base64 encoding for images
🚨 Error handling: Context window overflow detection

Next Steps

Update PR description (use template above as starting point)
Make dependency optional (follow OpenAI pattern)
Rename test file (anthropic.ts → anthropic.test.ts)
Investigate API key callbacks (check Anthropic SDK types)
Consider toBase64 approach (or document why chunking is needed)
Explain optional chaining change (or revert if unintentional)

Once these are addressed, this will be ready to merge! The core implementation is solid. 🚀

Questions?

Feel free to ask if you need clarification on any of these points. The maintainers and community are here to help! 👍

🦆

🤖 This is an experimental AI agent response from the Strands team, powered by Strands Agents. We're exploring how AI agents can help with community support and development. Your feedback helps us improve! If you'd prefer human assistance, please let us know.

Unshure · 2026-01-14T19:21:22Z

test/integ/anthropic.test.ts

+      expect(fullText.toLowerCase()).toContain('yellow')
+    })
+
+    it('processes PDF document input correctly', async () => {


nit: Can we move this to the agent.test.ts file? This should be tested across all model providers.

Unshure · 2026-01-14T19:22:07Z

test/integ/anthropic.test.ts

+    })
+  })
+
+  describe('Agent Tool Use', () => {


nit: We already test tools in agent.test.ts, we dont need the duplication here.

Unshure · 2026-01-14T19:22:54Z

test/integ/agent.test.ts

-
-// Import fixtures using Vite's ?url suffix


nit: Why remove this? I wouldnt know what this is without this comment

Unshure · 2026-01-14T19:36:49Z

src/models/anthropic.ts

+
+    const request: Anthropic.MessageStreamParams = {
+      model: this._config.modelId,
+      max_tokens: this._config.maxTokens ?? 1024,


nit: This should already be set. We have a default of 4096 above, which is honestly kinda low.

I would either require it as input (same as what we do in python), or create a map of different anthropic model ids to their max tokens. Seems like Haiku 3 is the only one that has a max of 4096, so we could also just do a check that if its haiku, we set it to 4096, otherwise we set it to 32k

https://platform.claude.com/docs/en/about-claude/models/overview

add support for anthropic

7b48b8f

chaynabors temporarily deployed to auto-approve December 29, 2025 22:09 — with GitHub Actions Inactive

minor fixes

b3c84c9

chaynabors had a problem deploying to auto-approve December 29, 2025 22:14 — with GitHub Actions Failure

remove useless comment

72b3b3d

chaynabors temporarily deployed to auto-approve December 29, 2025 22:15 — with GitHub Actions Inactive

repull lockfile

214e0b3

chaynabors had a problem deploying to auto-approve December 29, 2025 22:21 — with GitHub Actions Failure

chaynabors self-assigned this Dec 29, 2025

chaynabors temporarily deployed to auto-approve December 30, 2025 18:18 — with GitHub Actions Inactive

make anthropic dependency optional

9a17dbc

chaynabors temporarily deployed to auto-approve December 30, 2025 18:54 — with GitHub Actions Inactive

zastrowm approved these changes Dec 30, 2025

View reviewed changes

Unshure reviewed Dec 30, 2025

View reviewed changes

package.json Outdated Show resolved Hide resolved

test/integ/anthropic.test.ts Show resolved Hide resolved

zastrowm requested changes Dec 31, 2025

View reviewed changes

test/integ/anthropic.test.ts Show resolved Hide resolved

strands-agent reviewed Jan 13, 2026

View reviewed changes

strands-agent mentioned this pull request Jan 13, 2026

[V1] Models - Gemini #388

Open

chaynabors added 2 commits January 14, 2026 11:12

Merge remote-tracking branch 'origin/main' into chay/anthropic

0548140

address comments

47b3dbf

chaynabors had a problem deploying to auto-approve January 14, 2026 16:33 — with GitHub Actions Failure

fix integ

6f43e9b

chaynabors temporarily deployed to auto-approve January 14, 2026 17:16 — with GitHub Actions Inactive

Unshure approved these changes Jan 14, 2026

View reviewed changes

feat: Support Anthropic as a model provider #374

Are you sure you want to change the base?

feat: Support Anthropic as a model provider #374

Uh oh!

Conversation

chaynabors commented Dec 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Public API Changes

Prompt Caching behavior

Use Cases

Implementation Notes

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

strands-agent left a comment

Choose a reason for hiding this comment

Review Summary

📝 Action Items

1. Update PR Description ⭐ High Priority

Use Cases

3. Fix Test File Name ⭐ High Priority

4. API Key Callback Support 🔍 Needs Investigation

5. Use toBase64 for Browser Compatibility 🔍 Investigate

6. Optional Chaining Change in openai.ts 🤔 Clarification Needed

✅ What's Already Great

Next Steps

Questions?

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

chaynabors commented Dec 29, 2025 •

edited

Loading

5. Use `toBase64` for Browser Compatibility 🔍 Investigate

6. Optional Chaining Change in `openai.ts` 🤔 Clarification Needed