Schema | humane-proxy

Describes the environment variables required to run the server.

Name	Required	Description	Default
`PATH`	No	Environment path to ensure the humane-proxy CLI is executable	/app/.venv/bin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
`GROQ_API_KEY`	No	Groq API Key for Stage 3 Reasoning (LlamaGuard)
`OPENAI_API_KEY`	No	OpenAI API Key for Stage 3 Reasoning (OpenAI Moderation or Chat)
`HUMANE_PROXY_PORT`	No	Proxy port	8000
`HUMANE_PROXY_ADMIN_KEY`	No	Bearer token for securing the /admin endpoints
`HUMANE_PROXY_ENABLED_STAGES`	No	Active stages as a JSON array string (e.g., '[1, 2, 3]')	[1]
`HUMANE_PROXY_RISK_THRESHOLD`	No	Score threshold for criminal_intent escalation	0.7
`HUMANE_PROXY_STAGE1_CEILING`	No	Early exit after Stage 1 if score is below this threshold	0.3
`HUMANE_PROXY_STAGE2_CEILING`	No	Early exit after Stage 2 if score is below this threshold	0.4
`HUMANE_PROXY_STAGE3_TIMEOUT`	No	Stage 3 timeout in seconds	10
`HUMANE_PROXY_STAGE3_PROVIDER`	No	Stage 3 provider (auto, openai_moderation, llamaguard, openai_chat, none)	auto

Features and capabilities supported by this server

Capability	Details
`tools`	{ "listChanged": true }
`logging`	{}
`prompts`	{ "listChanged": false }
`resources`	{ "subscribe": false, "listChanged": false }
`experimental`	{}

Functions exposed to the LLM to take actions

Name	Description
check_message_safetyA	Classify a message for self-harm or criminal intent. Parameters message: The user message to classify. session_id: Optional session identifier for trajectory tracking. Returns dict `{"safe": bool, "category": str, "score": float, "triggers": list, "stage_reached": int, "should_escalate": bool, ...}`
get_session_riskA	Return the current risk trajectory for a session. Parameters session_id: The session identifier to query. Returns dict `{"spike_detected": bool, "trend": str, "window_scores": list, "category_counts": dict, "message_count": int}`
list_recent_escalationsA	Return recent escalation events from the audit log. Parameters limit: Maximum number of events to return (default 20). category: Filter by category (`"self_harm"` or `"criminal_intent"`). Omit for all categories. Returns list[dict] List of escalation records.

Interactive templates invoked by user choice

Name	Description
No prompts

Contextual data attached and managed by the client

Name	Description
No resources

humane-proxy