Anthropic SDK Integration

Wrappers for the official Anthropic Python SDK with THSP safety validation.

Installation

pip install sentinelseed[anthropic]

Components

Quick Start

Option 1: Drop-in Replacement

from sentinelseed.integrations.anthropic_sdk import SentinelAnthropic

client = SentinelAnthropic(api_key="...")

message = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}]
)

Option 2: Wrap Existing Client

from anthropic import Anthropic
from sentinelseed.integrations.anthropic_sdk import wrap_anthropic_client

client = Anthropic()
safe_client = wrap_anthropic_client(
    client,
    seed_level="standard",
    validate_input=True,
)

Option 3: Inject Seed Only

from anthropic import Anthropic
from sentinelseed.integrations.anthropic_sdk import inject_seed

client = Anthropic()
message = client.messages.create(
    model="claude-sonnet-4-20250514",
    system=inject_seed("You are a helpful assistant"),
    messages=[...]
)

Configuration

SentinelAnthropic(
    api_key=None,
    seed_level="standard",
    enable_seed_injection=True,
    validate_input=True,
    validate_output=True,
    validation_model="claude-3-5-haiku-20241022",
    use_heuristic_fallback=True,
    block_unsafe_output=False,
    fail_closed=False,
    validation_timeout=30.0,
    max_text_size=50*1024,
)

Validation Modes

Heuristic (Fast)

Content checked against local patterns first. No API calls.

Semantic (LLM-based)

Deep analysis by LLM for subtle violations.

Flow:

1. Heuristic validation (fast, free)

2. If passes, semantic validation (slower, API cost)

3. Both must pass

Blocked Response Format

{
    "id": "blocked",
    "type": "message",
    "role": "assistant",
    "content": [{"type": "text", "text": "Input blocked by Sentinel: [reason]"}],
    "model": "sentinel-blocked",
    "stop_reason": "sentinel_blocked",
    "sentinel_blocked": True,
    "sentinel_gate": "harm",
}

Links

• Anthropic Docs