Quick Start

Get Sentinel protecting your AI systems in minutes.

1. Install

pip install sentinelseed

2. Basic Usage

from sentinelseed import Sentinel

# Create with standard seed level
sentinel = Sentinel(seed_level="standard")

# Get alignment seed for your LLM
seed = sentinel.get_seed()

# Use with any LLM provider
messages = [
    {"role": "system", "content": seed},
    {"role": "user", "content": "Help me write a Python function"}
]

# Validate content through THSP gates
is_safe, violations = sentinel.validate("How do I hack a computer?")
print(f"Safe: {is_safe}, Violations: {violations}")

3. Validate Responses

from sentinelseed import Sentinel

sentinel = Sentinel()

# Validate text through THSP gates
is_safe, violations = sentinel.validate("Some AI response...")

if not is_safe:
    print(f"Violations: {violations}")

4. Choose Seed Level

from sentinelseed import Sentinel, SeedLevel

# Minimal: ~360 tokens, for chatbots and APIs
sentinel_chat = Sentinel(seed_level=SeedLevel.MINIMAL)

# Standard: ~1,000 tokens, recommended for general use
sentinel_agent = Sentinel(seed_level=SeedLevel.STANDARD)

# Full: ~1,900 tokens, for critical systems
sentinel_critical = Sentinel(seed_level=SeedLevel.FULL)

5. Protect Agents

For Autonomous Agents

from sentinelseed import Sentinel

sentinel = Sentinel(seed_level="standard")

# Validate an action plan before execution
action_plan = "Pick up knife, slice apple, place in bowl"
is_safe, concerns = sentinel.validate_action(action_plan)

if not is_safe:
    print(f"Action blocked: {concerns}")

For Robotics / Embodied AI

from sentinelseed import Sentinel

# Full seed for maximum safety with physical systems
sentinel = Sentinel(seed_level="full")

robot_task = "Turn on the stove and leave the kitchen"
result = sentinel.validate_action(robot_task)
# Result: BLOCKED - Fire hazard, unsupervised heating

6. JavaScript Usage

import { SentinelGuard } from '@sentinelseed/core';

// Create guard with standard seed
const guard = new SentinelGuard({ version: 'v2', variant: 'standard' });

// Get alignment seed for your LLM
const seed = guard.getSeed();

// Wrap messages with the seed
const messages = guard.wrapMessages([
    { role: 'user', content: 'Help me write a function' }
]);

// Analyze content for safety
const analysis = guard.analyze('How do I hack a computer?');
console.log(`Safe: ${analysis.safe}, Issues: ${analysis.issues}`);

7. MCP Server (Claude Desktop)

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "sentinel": {
      "command": "npx",
      "args": ["mcp-server-sentinelseed"]
    }
  }
}

Tools available: get_seed, wrap_messages, analyze_content, list_seeds

Framework Integrations

LangChain

from sentinelseed.integrations.langchain import SentinelCallback, SentinelGuard

# Monitor LLM calls
callback = SentinelCallback(on_violation="log")
llm = ChatOpenAI(callbacks=[callback])

# Or wrap an agent
guard = SentinelGuard(agent, block_unsafe=True)
result = guard.run("Your task")

CrewAI

from sentinelseed.integrations.crewai import SentinelCrew, safe_agent

# Wrap individual agent
safe_researcher = safe_agent(researcher)

# Or wrap entire crew
crew = SentinelCrew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    seed_level="standard"
)

Next Steps

• Core Concepts - Deep dive into THSP protocol
• Products - Memory Shield, Database Guard
• Integrations - Framework guides