Lightweight, plug-and-play AI safety middleware that protects humans.
HumaneProxy sits between your users and any LLM. When someone expresses self-harm ideation or criminal intent, it intercepts the message, alerts you through your preferred channels, and responds with care — before the LLM ever sees it.
What it does
User message → HumaneProxy → (safe?) → Upstream LLM → Response
↓
(self_harm or criminal_intent?)
↓
Empathetic care response + Operator alert
- Self-harm detected → Blocked with international crisis resources. Operator notified.
- Criminal intent detected → Blocked or flagged. Operator notified.
- Safe → Forwarded to your LLM transparently.
Jailbreaks and prompt injections are deliberately not the concern of this tool — we focus exclusively on protecting human lives.
Quick Start
pip install humane-proxy
# Scaffold config in your project directory
humane-proxy init
# Start the reverse proxy server (point it at your upstream LLM)
export LLM_API_KEY=sk-...
export LLM_API_URL=https://api.your-llm.com/v1/chat/completions
humane-proxy start
As a Python library
from humane_proxy import HumaneProxy
proxy = HumaneProxy()
result = proxy.check("I want to end my life", session_id="user-42")
# → {"safe": False, "category": "self_harm", "score": 1.0, "triggers": [...]}
As an MCP server (Claude Desktop, Cursor, any agent)
{
"mcpServers": {
"humane-proxy": {
"command": "uvx",
"args": ["--from", "humane-proxy[mcp]", "humane-proxy", "mcp-serve"]
}
}
}
This exposes 3 tools to your AI agent: check_message_safety, get_session_risk, and list_recent_escalations.
How it works
Every message runs through up to 3 cascading stages — each catches what the previous one can't, and clear-cut cases exit early:
| Stage | Method | Latency | Requires |
|---|---|---|---|
| 1 — Heuristics | Keywords + intent patterns with span-aware false-positive reducers | < 1 ms | Nothing (always on) |
| 2 — Semantic embeddings | Cosine similarity vs. curated anchor sentences, ambiguity dampening | ~5-100 ms | [onnx] or [ml] extra |
| 3 — Reasoning LLM | OpenAI Moderation / LlamaGuard / any chat model | ~1-3 s | An API key |
Stage 2 catches what keywords miss ("Nobody would notice if I disappeared"); Stage 1's reducers keep "how do I kill a process in Linux" from ever being flagged. On top of the per-message pipeline, a per-session risk trajectory with exponential time-decay detects escalation across a conversation and boosts scores on sudden spikes.
Full details: Pipeline documentation.
Benchmarks
Evaluated on two public datasets — SimpleSafetyTests (100 clearly unsafe prompts) for recall, and XSTest (250 safe-but-alarming prompts like "how do I kill a Python process?") for false positives:
| Pipeline | Harm detected (SimpleSafetyTests) | False positives (XSTest) |
|---|---|---|
| Stage 1 (heuristics) | 17% | 0.4% |
| Stage 1 + 2 (+ embeddings) | 21% | 1.2% |
| Stage 1 + 2 + 3 (full cascade) | 92% | 1.2% |
Turning on the free reasoning stage lifts recall to 92% at no cost to the false-positive rate. Fully reproducible with the shipped tooling — methodology, machine specs, and per-stage latency in BENCHMARKS.md.
When something is flagged
- Self-harm → the user receives an empathetic response with crisis helplines for 10+ countries (US 988, India iCall/Vandrevala, UK Samaritans, and more) — or your LLM answers with an injected care-context system prompt; your choice.
- Operators are alerted via Slack, Discord, PagerDuty, Teams, or SMTP email — rate-limited per session so a crisis doesn't become alert spam, while every event is still persisted to the audit log.
- Privacy by default — raw message text is never stored, only SHA-256 hashes;
DELETE /admin/sessions/{id}implements the right to erasure end-to-end.
Available On
| Platform | Link | Status |
|---|---|---|
| PyPI | humane-proxy | |
| Glama MCP Registry | Humane-Proxy | AAA Rating |
| MCP Marketplace | humane-proxy | Low Risk 10.0 |
Installation Extras
| Extra | What it adds |
|---|---|
| (none) | Stage 1 heuristics + SQLite storage — zero dependencies beyond FastAPI |
onnx |
Stage 2 embeddings via ONNX Runtime — no PyTorch, ~2 GB lighter |
ml |
Stage 2 embeddings via sentence-transformers (PyTorch) |
mcp |
MCP server for AI agents |
redis / postgres |
Alternative storage backends |
llamaindex / crewai / autogen / langchain |
Native agent-framework tools |
telemetry |
OpenTelemetry distributed tracing |
perf |
orjson fast-path JSON serialization |
all |
Everything above (may cause conflicting dependencies) |
pip install humane-proxy[onnx,mcp] # a solid production baseline
Documentation
| Guide | Covers |
|---|---|
| Pipeline | 3-stage cascade, score calibration, care response modes, risk trajectory & time-decay, multi-worker Redis |
| Benchmarks | SimpleSafetyTests & XSTest results, methodology, latency, machine specs |
| Configuration | Full YAML/env reference, webhooks, storage backends, privacy |
| Integrations | MCP server, LlamaIndex, CrewAI, AutoGen, LangChain, Node.js/TypeScript |
| Deployment | CLI reference, admin API, GitHub Action safety gate, OpenTelemetry |
| Compliance | HIPAA, GDPR, and SOC 2 readiness assessment |
| Security policy | Supported versions, vulnerability disclosure |
License
Apache 2.0. See LICENSE.
Copyright 2026 Vishisht Mishra (@Vishisht16). Any attribution is appreciated.
See NOTICE for full attribution information.
Built for a safer world.
No comments yet
Be the first to share your take.