agentforge-anthropic 0.3.0

Anthropic native LLM provider for AgentForge — Claude with caching, extended thinking, streaming

agentforge-anthropic

Anthropic native LLM provider for AgentForge.

Use this package when you want to call Claude directly via Anthropic's native API rather than through AWS Bedrock. Native access ships latest-model support, prompt caching, extended thinking, and per-token streaming the day Anthropic releases them.

Install

pip install "agentforge-anthropic[anthropic]"

The bare pip install agentforge-anthropic install keeps the package importable for tests; the production runner needs the anthropic SDK, which the [anthropic] extra pulls in.

Use

from agentforge import Agent
from agentforge_anthropic import AnthropicClient

# Direct instantiation
client = AnthropicClient.from_config(
    model="claude-sonnet-4-7",
    api_key="sk-ant-...",  # or omit and use ANTHROPIC_API_KEY
)

# Or via Agent's resolver (recommended)
agent = Agent(model="anthropic:claude-sonnet-4-7")

Capabilities

AnthropicClient.capabilities() declares the closed-vocabulary set:

• tools — Claude's tool-use API
• json_mode — JSON response format via system-prompt addendum
• caching — prompt caching with cache_control: ephemeral blocks
• thinking — extended thinking via thinking={"type": "enabled", "budget_tokens": ...}
• streaming — per-token streaming via messages.stream()

Callers gate optional methods on client.supports("capability") before invoking — the additive contract in ADR-0009.

Why a separate sister package from agentforge-bedrock?

Both implement the same LLMClient contract but talk to different endpoints:

• agentforge-bedrock — Bedrock Converse API; Anthropic models routed via AWS; cross-region inference profiles; AWS IAM auth.
• agentforge-anthropic — Anthropic native API; direct; bearer-token auth; faster access to new model releases.

A string-id swap ("bedrock:..." → "anthropic:...") is the only change in caller code.

Testing

Production runner under # pragma: no cover wraps the anthropic SDK. Unit tests inject a FakeAnthropicRunner (in agentforge_anthropic._inmem_runner) that records every messages_create call. Live integration tests are gated behind pytest -m live and need a real ANTHROPIC_API_KEY.

License

Apache-2.0. See LICENSE.