User message → HumaneProxy → (safe?) → Upstream LLM → Response
                    ↓
              (self_harm or criminal_intent?)
                    ↓
              Empathetic care response  +  Operator alert

pip install humane-proxy

# Scaffold config in your project directory
humane-proxy init

# Start the reverse proxy server
# (requires LLM_API_KEY and LLM_API_URL in .env — these point to your upstream LLM)
humane-proxy start

from humane_proxy import HumaneProxy

proxy = HumaneProxy()

# Sync check (Stages 1+2)
result = proxy.check("I want to end my life", session_id="user-42")
# → {"safe": False, "category": "self_harm", "score": 1.0, "triggers": [...]}

# Async check (all 3 stages)
result = await proxy.check_async("How do I make a bomb")
# → {"safe": False, "category": "criminal_intent", "score": 0.9, ...}

pip install humane-proxy[mcp]

# Start the MCP server (stdio transport — for Claude Desktop, Cursor, etc.)
humane-proxy mcp-serve

{
  "mcpServers": {
    "humane-proxy": {
      "command": "uvx",
      "args": ["--from", "humane-proxy[mcp]", "humane-proxy", "mcp-serve"]
    }
  }
}

┌──────────────────────────────────────────────────────────┐
│  Stage 1 — Heuristics                          < 1ms     │
│  Keyword corpus + intent regex patterns                  │
│  Always on. Catches clear cases instantly.               │
│  Early-exit: definitive self_harm → block immediately.   │
└──────────────────────────────────────────────────────────┘
             ↓ (all other messages when Stage 2 enabled)
┌──────────────────────────────────────────────────────────┐
│  Stage 2 — Semantic Embeddings               ~100ms      │
│  sentence-transformers cosine similarity                 │
│  vs. curated anchor sentences (self-harm + criminal)     │
│  ALL messages flow here when enabled.                    │
│  Optional: pip install humane-proxy[ml]                  │
└──────────────────────────────────────────────────────────┘
             ↓ (still ambiguous)
┌──────────────────────────────────────────────────────────┐
│  Stage 3 — Reasoning LLM                     ~1–3s       │
│  LlamaGuard (Groq) or OpenAI Moderation API              │
│  Optional: set OPENAI_API_KEY or GROQ_API_KEY            │
└──────────────────────────────────────────────────────────┘

pipeline:
  # Which stages to run. [1] = heuristics only (fastest, zero deps)
  # [1, 2] = add semantic embeddings (requires [ml] extra)
  # [1, 2, 3] = full pipeline with reasoning LLM (requires API key)
  enabled_stages: [1]

  # Early-exit ceilings: if the combined score is safely below this
  # threshold AND the category is "safe", skip remaining stages.
  stage1_ceiling: 0.3    # exit after Stage 1 if score ≤ 0.3 and safe
  stage2_ceiling: 0.4    # exit after Stage 2 if score ≤ 0.4 and safe

pip install humane-proxy[ml]

pipeline:
  enabled_stages: [1, 2]

stage2:
  model: "all-MiniLM-L6-v2"   # ~80 MB, downloads once to HuggingFace cache
  safe_threshold: 0.35         # cosine similarity below this → safe

# Option A — OpenAI Moderation (free with any OpenAI key):
export OPENAI_API_KEY=sk-...

# Option B — LlamaGuard via Groq (free tier, very fast):
export GROQ_API_KEY=gsk_...

pipeline:
  enabled_stages: [1, 2, 3]

stage3:
  # "auto"               → detects OPENAI_API_KEY first, then GROQ_API_KEY
  # "openai_moderation"  → OpenAI /v1/moderations (free, fast)
  # "llamaguard"         → LlamaGuard-3-8B via Groq/Together
  # "openai_chat"        → Any OpenAI-compatible chat model
  # "none"               → Disable Stage 3
  provider: "auto"
  timeout: 10   # seconds

  openai_moderation:
    api_url: "https://api.openai.com/v1/moderations"

  llamaguard:
    api_url: "https://api.groq.com/openai/v1/chat/completions"
    model: "meta-llama/llama-guard-3-8b"

  openai_chat:
    api_url: "https://api.openai.com/v1/chat/completions"
    model: "gpt-4o-mini"

safety:
  categories:
    self_harm:
      # Self-harm escalation threshold (0.0 to 1.0).
      # Scores below this are downgraded to safe.
      escalate_threshold: 0.5

      response_mode: "block"     # default

      # Optional: override the built-in message
      block_message: "We're here for you. Please reach out to..."

safety:
  categories:
    self_harm:
      response_mode: "forward"

delta = current_score − weighted_mean(last N scores)
spike = delta > 0.35    (configurable via spike_delta)

trajectory:
  window_size: 5          # messages in rolling window
  spike_delta: 0.35       # delta threshold for spike detection

  # Half-life in hours.  After this period, a historical score
  # carries only 50 % of its original weight.
  #   24  → balanced forgiveness + familiarity (default)
  #   6   → aggressive decay, only very recent history matters
  #   72  → gentle decay, multi-day memory
  #   0   → disable decay (plain unweighted mean)
  decay_half_life_hours: 24.0

export HUMANE_PROXY_DECAY_HALF_LIFE=12   # 12-hour half-life

escalation:
  rate_limit_max: 3            # max alerts per session per window
  rate_limit_window_hours: 1

  webhooks:
    slack_url: "https://hooks.slack.com/services/..."
    discord_url: "https://discord.com/api/webhooks/..."
    pagerduty_routing_key: "your-routing-key"
    teams_url: "https://outlook.office.com/webhook/..."

    # Email alerts via SMTP (stdlib, no extra deps)
    email:
      host: "smtp.gmail.com"
      port: 587
      use_tls: true
      username: "your@gmail.com"
      password: "app-password"
      from: "humane-proxy@yourorg.com"
      to:
        - "safety-team@yourorg.com"
        - "oncall@yourorg.com"

# Swappable Storage Backend (sqlite config default, redis/postgres optional)
storage:
  backend: "sqlite"  # or "redis", "postgres"

# Safety check
hp check "I want to end my life"
# 🆘 FLAGGED — self_harm
# Score   : 1.0
# Category: self_harm

# Run benchmark evaluation
hp benchmark --dataset evals/sample.json
hp benchmark --dataset evals/sample.json --ci  # exit code 1 on failure

# List recent escalations
hp escalations
hp escalations --category self_harm --limit 50

# Session risk history
hp session user-42

# Start proxy server
hp start [--host 0.0.0.0] [--port 8000]

# MCP server (requires [mcp] extra)
hp mcp-serve

# .github/workflows/safety-benchmark.yml
name: Safety Benchmark
on: [push, pull_request]

jobs:
  benchmark:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: Vishisht16/Humane-Proxy@v0.4.0
        with:
          dataset: evals/sample.json

export HUMANE_PROXY_ADMIN_KEY=your-secret-key

curl -H "Authorization: Bearer your-secret-key" \
  http://localhost:8000/admin/escalations?category=self_harm&limit=10

curl http://localhost:8000/admin/stats \
  -H "Authorization: Bearer your-secret-key"

# Delete session data (right to erasure)
curl -X DELETE http://localhost:8000/admin/sessions/user-42 \
  -H "Authorization: Bearer your-secret-key"

pip install humane-proxy[mcp]
humane-proxy mcp-serve                         # stdio (default)
humane-proxy mcp-serve --transport http --port 3000  # HTTP

pip install humane-proxy[llamaindex]

from humane_proxy.integrations.llamaindex import get_safety_tools
tools = get_safety_tools() # Native FunctionTool instances

pip install humane-proxy[crewai]

from humane_proxy.integrations.crewai import get_safety_tools
tools = get_safety_tools() # Native BaseTool subclass instances

pip install humane-proxy[autogen]

from humane_proxy.integrations.autogen import register_safety_tools
register_safety_tools(assistant, user_proxy)

pip install humane-proxy[langchain]

from humane_proxy.integrations.langchain import get_safety_tools

# Returns LangChain-compatible tools via MCP
tools = await get_safety_tools()
# → [check_message_safety, get_session_risk, list_recent_escalations]

# Or get the config dict for MultiServerMCPClient:
from humane_proxy.integrations.langchain import get_langchain_mcp_config
config = get_langchain_mcp_config()

privacy:
  store_message_text: true