Anthropic Plugin Deep Dive

Overview

The Anthropic plugin connects Ever Works to Anthropic's Claude API for AI-powered content generation and conversational AI. It extends the BaseAiProvider abstract class and uses the shared AiOperations utility (wrapping LangChain) to provide chat completions and streaming responses. Claude models are recognized for their large context windows and strong instruction-following capabilities, making them well-suited for work content generation.

Architecture

The plugin extends BaseAiProvider from @ever-works/plugin/abstract and delegates all API communication to AiOperations from @ever-works/plugin/ai, which wraps LangChain's Anthropic integration.

AnthropicPlugin (extends BaseAiProvider)
  |-- AiOperations (@ever-works/plugin/ai)
       |-- LangChain (@langchain/anthropic)
            |-- Anthropic Messages API

On onLoad, the plugin creates an AiOperations instance configured with providerType: 'anthropic' and the default Claude Sonnet model. Each request resolves its own configuration from the settings hierarchy, enabling per-request model and key customisation.

Configuration

Environment Variables

Variable	Required	Description
N/A	--	No environment-variable fallbacks; users provide their own API key

Settings Schema

interface AnthropicSettings {
	apiKey: string; // Anthropic API key (x-secret, user-scoped, required)
	defaultModel: string; // Default model (default: 'claude-sonnet-4-5-20250514')
	simpleModel: string; // Simple tasks model (default: 'claude-haiku-4-5-20251001')
	mediumModel: string; // Standard tasks model (default: 'claude-sonnet-4-5-20250929')
	complexModel: string; // Complex tasks model (default: 'claude-sonnet-4-5-20250514')
	baseUrl: string; // Custom API endpoint (default: 'https://api.anthropic.com/v1/', hidden)
	temperature: number; // Response randomness, 0-2 (default: 0.7, hidden)
	maxTokens: number; // Max response length (default: 4096, hidden)
}

• configurationMode: user-required -- each user must provide their own Anthropic API key.
• Model fields use the x-widget: 'model-select' custom widget.
• The baseUrl setting supports custom proxies or compatible API services.

Capabilities

Capability	Supported	Details
Structured output	Yes	JSON mode and tool use
Streaming	Yes	Server-sent events via async iterables
Tool calling	Yes	Tool use for structured data extraction
Vision	Yes	Image analysis with Claude models
Embeddings	No	Throws Error('Embeddings not supported by Anthropic')
Max context	200,000 tokens	Claude's extended context window

Important: Anthropic does not offer an embeddings API. The createEmbedding method throws an explicit error. If you need embeddings alongside Anthropic for chat, use a secondary provider (e.g., OpenAI) for embedding generation.

API Reference

Chat Completion

Method	Signature	Description
createChatCompletion	(options: ChatCompletionOptions) => Promise<ChatCompletionResponse>	Single-shot completion
createStreamingChatCompletion	(options: ChatCompletionOptions) => AsyncIterable<ChatCompletionChunk>	Streaming completion via async generator

Embeddings

Method	Signature	Description
createEmbedding	(options: EmbeddingOptions) => Promise<EmbeddingResponse>	Not supported -- always throws

Model Management

Method	Signature	Description
listModels	(settings?) => Promise<readonly AiModel[]>	Lists available Claude models
isAvailable	(settings?) => Promise<boolean>	Tests connection to Anthropic API
getCapabilities	() => AiModelCapabilities	Returns static capability flags

Implementation Details

Tiered Model Resolution

The inherited resolveConfig() method resolves the appropriate Claude model for the task tier:

• Simple tier: claude-haiku-4-5-20251001 -- optimized for speed; handles tags, short descriptions, quick classifications.
• Medium tier: claude-sonnet-4-5-20250929 -- balanced speed and quality; handles listings, summaries, content reformatting.
• Complex tier: claude-sonnet-4-5-20250514 -- highest quality; handles full page generation and multi-step analysis.
• Default: claude-sonnet-4-5-20250514 -- fallback for unspecified tiers.

No Embedding Support

Unlike OpenAI, Anthropic does not provide embedding models. The createEmbedding method is implemented as a hard error:

async createEmbedding(_options: EmbeddingOptions): Promise<EmbeddingResponse> {
  throw new Error('Embeddings not supported by Anthropic');
}

Works that require semantic search should pair Anthropic (for generation) with a separate embedding provider.

200K Context Window

Claude models support up to 200,000 tokens of context, which is reported via getCapabilities(). This allows processing large source documents or entire website contents in a single request during work generation.

Usage Examples

// Non-streaming chat completion with Claude
const response = await anthropicPlugin.createChatCompletion({
	messages: [
		{ role: 'system', content: 'You are a work content writer.' },
		{ role: 'user', content: 'Write a detailed description for Acme Corp.' }
	],
	settings: { apiKey: userApiKey, defaultModel: 'claude-sonnet-4-5-20250514' }
});

// Streaming chat completion
for await (const chunk of anthropicPlugin.createStreamingChatCompletion({
	messages: [{ role: 'user', content: 'Summarize this company...' }],
	settings: { apiKey: userApiKey }
})) {
	process.stdout.write(chunk.content || '');
}

// List available models
const models = await anthropicPlugin.listModels({ apiKey: userApiKey });

// Check availability
const available = await anthropicPlugin.isAvailable({ apiKey: userApiKey });

Rate Limiting & Quotas

• Anthropic API limits: Depend on the user's usage tier. Tier 1 starts at 50 RPM and 40,000 TPM; higher tiers increase these limits significantly.
• Max tokens: Defaults to 4,096 per response. Claude models support up to 8,192 output tokens; configurable via the maxTokens setting.
• Context window: Up to 200,000 tokens input.
• The plugin does not implement internal rate-limit tracking. Anthropic returns 429 errors which propagate to callers.

Error Handling

• Plugin not loaded: All methods throw Error('Anthropic plugin not loaded') if called before onLoad.
• Embedding requests: Always throw Error('Embeddings not supported by Anthropic').
• Connection test: isAvailable calls testConnection and returns false on any failure, including invalid API keys or network errors.
• LangChain errors: Errors from the underlying LangChain Anthropic SDK propagate directly. Common errors include authentication failure (401), rate limiting (429), and overloaded API (529).

Related Plugins

• OpenAI Plugin Deep Dive -- alternative AI provider with embedding support.
• Ollama Plugin Deep Dive -- self-hosted AI provider for private inference.
• Anthropic Plugin -- overview documentation for the Anthropic plugin.