@veryfront/ext-llm-anthropic

Category: LLM | Contract: LLMProvider | Built-in

Provides Anthropic Claude models for Veryfront agents and chat. Once loaded, any model string prefixed with anthropic/ (e.g. anthropic/claude-sonnet-4-6) is routed through this provider via the LLMProviderRegistry.

Configuration

The extension is configured through the standard LLMProviderConfig interface:

| Field | Required | Description | | ------------ | -------- | ----------------------------------------------------------------------------------- | | credential | Yes | Anthropic API key (maps to x-api-key header) | | baseURL | No | Override the messages endpoint base (default: https://api.anthropic.com/v1) | | authToken | No | Bearer token for proxied deployments (Veryfront Cloud, Bedrock-compatible gateways) | | name | No | Display name for errors and telemetry (default: "anthropic") | | fetch | No | Custom fetch implementation (used by veryfront-cloud for injecting project auth) |

Example

import extAnthropic from "@veryfront/ext-llm-anthropic";

const ext = extAnthropic();
ext.setup({
  require: (name) => registry, // LLMProviderRegistry
  // ...
});

// Then use via the registry:
const provider = registry.require("anthropic");
const runtime = provider.createModel("claude-sonnet-4-6", {
  credential: process.env.ANTHROPIC_API_KEY!,
  baseURL: "https://api.anthropic.com/v1",
});

Supported Features

Models

Automatic max_tokens defaults based on model family:

| Model | Default max_tokens | | ---------------------------- | ------------------ | | Claude Opus/Sonnet 4.6 | 128,000 | | Claude Opus/Sonnet/Haiku 4.5 | 64,000 | | Claude Opus 4.1 | 32,000 | | Claude 3 Haiku | 4,096 | | Unknown models | 4,096 |

Caller-provided maxOutputTokens is clamped at the model ceiling for known models.

Extended Thinking

Enable reasoning via the reasoning option:

runtime.doGenerate({
  prompt: [...],
  reasoning: {
    enabled: true,
    effort: "high",    // "low" | "medium" | "high" | "max"
    budgetTokens: 8192 // optional explicit override
  },
});

Effort-to-budget mapping: low = 1024, medium = 4096, high = 16384, max = 32768.

When thinking is enabled, temperature and topP are automatically dropped (Anthropic rejects the combo).

Prompt Caching

Control cache breakpoints via cacheControl:

runtime.doGenerate({
  prompt: [...],
  tools: [...],
  cacheControl: {
    system: true,   // or "5m" | "1h" | false
    tools: "1h",    // breakpoint on the last tool entry
  },
});

Provider Tools

Anthropic-native tools are supported via type: "provider" tool definitions:

| Short ID | Resolved Type | | ----------------------------------------------- | ------------------------- | | anthropic.code_execution | code_execution_20260120 | | anthropic.computer_use / anthropic.computer | computer_20250124 | | anthropic.text_editor | text_editor_20250728 | | anthropic.bash | bash_20250124 | | anthropic.memory | memory_20250818 | | anthropic.web_search | web_search_20250305 | | anthropic.web_fetch | web_fetch_20250910 |

Already-versioned IDs (e.g. anthropic.code_execution_20250522) pass through verbatim.

MCP Servers

Pass Anthropic-native MCP servers via mcpServers. Keys are automatically converted from camelCase to snake_case:

runtime.doGenerate({
  prompt: [...],
  mcpServers: [{
    type: "url",
    url: "https://example.com/mcp",
    name: "my-server",
    authorizationToken: "Bearer ...",
    toolConfiguration: {
      enabled: true,
      allowedTools: ["search"],
    },
  }],
});

Container

Pass anthropicContainer to attach a container context to the request (for computer-use sessions).

Provider Options

Arbitrary Anthropic-specific fields can be merged into the request body via providerOptions:

runtime.doGenerate({
  prompt: [...],
  providerOptions: {
    anthropic: { top_k: 3 },
    "my-custom-name": { metadata: { trace: "yes" } },
  },
});

Both the "anthropic" key and the provider's custom name are merged.

Unsupported Options (emits warnings)

The following unified options have no Anthropic equivalent and are silently dropped with a warning:

• presencePenalty
• frequencyPenalty
• seed
• topK
• responseFormat (non-text; use tool schemas instead)
• stopSequences beyond 4 entries (extras truncated)

Running Tests

deno task test