Anthropic¶
Claude implements the Anthropic Messages API. It translates shared completion settings to Anthropic payloads and returns the same Completion type as the other adapters. Embeddings and decisions aren't supported.
Connection¶
import intuit;
auto ep = new Claude("https://api.anthropic.com", "your-api-key");
auto cfg = cast(ClaudeModelConfig)ep.config("claude-sonnet-4-5");
cfg.maxTokens = 2048;
The constructor accepts a base url, optional key, and display name (default "Claude"). The endpoint sends a nonempty key in x-api-key and sets anthropic-version to 2023-06-01.
Use the host without /v1 or a trailing slash. Discovery requests GET /v1/models, and completions use POST /v1/messages. Discovery reads the returned data array without following pagination.
Set maxTokens before requesting a completion when the server requires it. Intuit leaves it unset by default.
Configuration¶
config(modelName) returns a cached ClaudeModelConfig. It supports the common temperature, topP, and maxTokens fields plus these settings.
| Field | Behavior |
|---|---|
topK |
Sends top_k when nonnegative. |
stopSequences |
Sends nonempty Anthropic stop sequences. Use this instead of the base stop field. |
system |
Prepends a system prompt to string system messages from the context. |
thinkingBudget |
Sends enabled thinking with this budget when nonnegative. |
The Messages completion builder doesn't serialize the base seed, stop, or params fields. Use the supported fields rather than assuming all base configuration settings apply.
Payload Translation¶
String system messages are removed from the message array and joined with newlines into the top-level system field. Non-string system content isn't included.
setResponseSchema places the normalized schema under output_config.format. Tool-choice values map to Anthropic's auto, any, and none forms, and a required tool uses its name. Tool definitions use input_schema.
Other message objects are passed through after system extraction. In particular, the adapter doesn't translate shared OpenAI-style assistant tool calls and tool-result messages into Anthropic request content blocks. Tool registration and response parsing work, but the shared tool round-trip examples aren't a complete Anthropic conversation adapter.
Responses¶
Text blocks populate Choice.text. Thinking blocks populate Choice.reasoning, while redacted thinking becomes [redacted thinking]. Tool-use blocks become shared ToolCall values.
The parser returns one choice and preserves the full response as both Completion.raw and Choice.raw. Prompt usage includes ordinary input, cache reads, and cache creation. Cache reads are reported as hits.
See Models for structured output, Responses for shared fields, and Errors for failure handling.