Humane Proxy
Put a safety layer in front of LLM traffic so agents flag self-harm and criminal-intent prompts before they reach your model or users.
Overview
Humane Proxy is an MCP server for the Ship phase that detects self-harm and criminal intent in LLM prompts as AI safety middleware.
What is this MCP server?
- Middleware that inspects LLM prompts for self-harm and criminal intent signals
- PyPI package humane-proxy with mcp-serve stdio entry
- Optional Stage 3 reasoning via OPENAI_API_KEY
- Optional LlamaGuard Stage 3 via GROQ_API_KEY
- Version 0.4.0 MCP server for agent-side safety registration
- Version 0.4.0 on PyPI with stdio transport via mcp-serve
- Two optional secret env vars: OPENAI_API_KEY and GROQ_API_KEY
- Described as AI safety middleware for self-harm and criminal intent detection
Community signal: 30 GitHub stars.
What problem does it solve?
Shipping a public agent without prompt-level safety screening exposes you to harmful use cases you did not design for.
Who is it for?
Builders launching user-facing chat agents who need a plug-in safety gate with optional OpenAI or Groq-backed stages.
Skip if: Internal-only codegen assistants with no end-user prompts or teams that already run enterprise DLP and moderation suites end-to-end.
What do I get? / Deliverables
Risky intent gets flagged in the MCP layer so you can block, log, or escalate before the main model responds.
- Prompt-level safety assessment hooks callable from your agent
- Configurable optional cloud reasoning stages
- Middleware integration point before main LLM calls
Recommended MCP Servers
Journey fit
How it compares
MCP safety middleware for prompts, not a full trust-and-safety operations platform or red-team skill.
Common Questions / FAQ
Who is Humane Proxy for?
Solo builders and small teams shipping LLM-powered products who need MCP-registered checks for self-harm and criminal intent without building classifiers from scratch.
When should I use Humane Proxy?
Enable it before public launch or beta when user prompts are unconstrained and you want agent-accessible safety screening in the inference path.
How do I add Humane Proxy to my agent?
Install the PyPI humane-proxy package, configure MCP stdio with the mcp-serve command, and optionally set OPENAI_API_KEY or GROQ_API_KEY for Stage 3 reasoning.