keel-llm-adapter-anthropic 0.1.2

Reference keel-llm-protocol adapter for Anthropic's Messages API — the divergent-provider grounding for the standard.

keel-llm-adapter-anthropic

A reference keel-llm-protocol adapter for Anthropic's Messages API (Claude).

This package is the divergent-provider grounding for the protocol standard. Where keel-llm-adapter-openai proves the standard against the OpenAI-compatible family, this proves it against an API that diverges on purpose — and the protocol absorbs every difference in the adapter, without changing the standard. That is the standard generalizing.

Is this for you?

Adopt when — you're consuming via keel-llm-protocol / keel-llm-reliability and want Claude calls mapped to the typed taxonomy (with Anthropic's divergences absorbed at the adapter layer). Skip when — you're calling Anthropic directly in a simple app — the official anthropic SDK has a wider surface (batches, files, prompt caching…); this adapter covers chat / streaming / tools.

Install

pip install keel-llm-adapter-anthropic     # pulls in keel-llm-protocol + httpx

Use

from keel_llm_adapter_anthropic import AnthropicAdapter
from keel_llm_protocol import system, user

adapter = AnthropicAdapter(model="claude-sonnet-4-5", api_key="...", provider="anthropic")
resp = await adapter.generate([system("Be terse."), user("What is a circuit breaker?")])
print(resp.text, resp.usage.total_tokens, resp.finish_reason)

Same keel-llm-protocol types as every other adapter — swap AnthropicAdapter for OpenAIAdapter and your code is unchanged. It implements ModelAdapter, StreamingModelAdapter, and ToolCallingModelAdapter.

How Anthropic diverges — and where it's handled

Every divergence below is absorbed inside the adapter; the protocol's types are untouched:

Anthropic divergence	Protocol stays the same because…
System prompt is a top-level system field, not a message	the adapter lifts Message(role="system") into the system param
Only user / assistant roles exist	tool messages → tool_result content blocks; assistant tool_calls → tool_use blocks
Tool input arrives as a parsed dict	the adapter json.dumps it into ToolCall.arguments (a JSON string)
Tools use input_schema + tool_choice: {type: auto|any|tool}	the adapter maps ToolSpec / tool_choice accordingly
max_tokens is required	the adapter supplies default_max_tokens (4096) when unset
Distinct SSE events (message_start / content_block_delta / message_delta)	the adapter maps them onto StreamChunk
529 overloaded status	mapped to retryable TransientError
{"type":"error","error":{...}} envelope + stop_reason values	mapped to the standard error taxonomy + FinishReason

Known limitation

Anthropic has no native response_format parameter — structured output is done via tools or prompt instructions. This adapter therefore treats ResponseFormat as a best-effort hint and does not enforce it. (A finding worth keeping: in the standard, response_format is a request, not a guarantee — provider support varies.)

Status

0.1.2 — 0.x while the API stabilizes through year one (breaking changes possible at minor bumps, documented in the CHANGELOG; pin exact versions).

The Keel toolkit

Composable, vendor-neutral LLM reliability libraries on PyPI: keel-llm-reliability · keel-llm-protocol · keel-llm-adapter-openai · keel-llm-adapter-anthropic · keel-llm-adapter-google · keel-circuit-breaker

MIT licensed.