Anthropic

@dsh-ai-providers/anthropic is the native Anthropic provider maintained by Firefox2100 as part of this project. It uses the Messages API directly rather than translating requests through an OpenAI-compatible endpoint. It is not an official DSH or Anthropic package.

Supported capability

Capability	Native endpoint
Language models	POST /messages
Model discovery	GET /models

Anthropic does not provide embedding, reranking, TTS, STT, or image-generation APIs, so this package does not register those services.

Configure a connection

Install the package and open Settings → Anthropic connections. The default connection uses:

Setting	Default
Base URL	https://api.anthropic.com/v1
Credential reference	ANTHROPIC_API_KEY
API version	2023-06-01
Stream silence limit	5 minutes

The secret is stored through DSH's credential service. Native requests authenticate with x-api-key and carry the configured anthropic-version header. Additional route headers can be used for beta features when required.

Configure routes and models

Create a named route, select a connection, and add exact Anthropic model IDs. The provider can discover accessible models from /models, or IDs can be entered manually. Each model can declare its context window, output limit, vision support, and reasoning levels.

Reasoning levels are native request fragments. For example, a model supporting adaptive thinking can use a level body such as:

{
  "thinking": {"type": "adaptive"},
  "output_config": {"effort": "high"}
}

Keep these fields aligned with the selected model. Anthropic changes thinking controls between model generations.

Native request mapping

The provider supports:

• top-level system prompts;
• text and image content;
• native tool declarations, tool_use, and tool_result blocks;
• streamed text, thinking, and tool input;
• prompt-cache read and write usage;
• temperature, top_p, top_k, stop sequences, output limits, and tool choice;
• additional native request fields through route or model configuration.

Files remain tool-readable text handles. Offloaded images remain placeholders. Non-offloaded images are read from DSH's attachment service and sent as base64 sources.

Errors and cancellation

HTTP and stream failures are mapped to DSH language-model failure classes, including authentication, rate limiting, overload, invalid requests, context overflow, timeout, and cancellation. retry-after and Anthropic request IDs are retained when the response provides them.