Humane-Proxy
MCP ServerVishisht16/Humane-Proxy
Humane-Proxy is a lightweight AI safety middleware designed to intercept and filter user messages for self-harm ideation or criminal intent before they reach an upstream LLM. It combines heuristic keyword matching, semantic embeddings, and reasoning models to provide real-time risk evaluation, empathetic user care, and operator alerts.
29 30Apache-2.0Updated 2026-07-29
89
GreatOverview
Humane-Proxy acts as a safety gatekeeper sitting between users and large language models, specifically focusing on safeguarding human well-being by catching self-harm and criminal intent signals. The middleware utilizes a multi-stage evaluation pipeline—ranging from instant regex heuristics and ONNX-based semantic embeddings to full reasoning LLM checks—alongside conversation-wide risk trajectory tracking. It supports flexible deployment models, including a standalone reverse proxy server, a Python library, an MCP server for AI agents, and integrations with major orchestration frameworks like LangChain and LlamaIndex.
Capabilities
- ▸Three-stage cascading safety pipeline (heuristics, semantic embeddings, reasoning LLM)
- ▸Session-level risk trajectory tracking with exponential time-decay
- ▸Automated operator alerting via Slack, Discord, PagerDuty, Teams, or SMTP
- ▸Model Context Protocol (MCP) server support exposing safety tools to AI agents
- ▸Privacy-first design storing message hashes instead of raw text with right-to-erasure support
- ▸Native integration tools for popular frameworks such as LangChain, LlamaIndex, CrewAI, and AutoGen
Best for
Intercepting self-harm text in chatbot applications to provide crisis support resources, Detecting criminal intent in user prompts and alerting administrative operators via webhook or email, Serving as an MCP server to add safety checks directly into Claude Desktop or other agentic workflows, Auditing and filtering requests in LLM-backed applications while maintaining data privacy via SHA-256 hashing
Works with
CursorClaude CodeChatGPTGitHub Copilot
Evoa Score breakdown= Σ (score × weight)
89Task usefulness20%100 → +20.0
実タスクにどれだけ役立つか（機能の豊富さ・用途の明確さ）。 — AIによるcapabilities/use-cases解析
Code quality15%92 → +13.8
実装・指示の品質。 — AIによるSKILL.md/README解析
Maintenance15%100 → +15.0
リポジトリがどれだけ活発に保守されているか。 — GitHub 最終push日時の新しさ
Documentation12%63 → +7.6
ドキュメントの充実度・分かりやすさ。 — README/独自要約の情報量
Security15%100 → +15.0
危険・不審な挙動が無いか。 — AIによるセキュリティレビュー
Originality10%88 → +8.8
ありふれたラッパーではない独自性。 — AIによる独自性判定
Popularity8%49 → +3.9
コミュニティの採用度。 — GitHub Stars/Forks（対数スケール）
Compatibility5%100 → +5.0
対応AIエージェントの広さ。 — AIによる対応エージェント判定
Weighted total89.1 / 100
ライセンス不明/制限あり（Red）のSkillは総合スコアに0.85倍の補正を適用します。 ランキングはこのScoreのみで決まり、広告で変わりません。 算出方法の詳細 →
Security considerations
Prioritizes user privacy by storing message hashes rather than raw conversation content and providing admin tools for complete data erasure.
Categories
Summary and analysis are original content generated by AI Skills Rank. The skill's source text is not reproduced here — view it on the linked repository.