How do I use Care Membrane MCP Server?

1. Click on "Install Server". 2. Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state. 3. In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Care Membrane MCP Server check if this response is ethically aligned: "I can't assist with that request as it violates safety guidelines."" That's it! The server will respond to your query, and you can continue using it as needed. Here is a step-by-step guide with screenshots.

care-membrane-mcp

Why this exists

AI safety today mostly means 'the model refuses certain prompts'. That's necessary but insufficient — it doesn't address the substrate problem: how does an AI agent know whether an action it's about to take serves or harms the people it's working with?

The Care Membrane is MEOK's reference implementation of a substrate-level care-validation layer. Before an agent commits to an action (sending an email, modifying a file, posting publicly, executing a transaction), the membrane evaluates the proposed action against four care dimensions: relational integrity, consent, dignity, and reversibility. Actions that fail care-validation are blocked or escalated to a human.

This isn't an alternative to model-level safety — it's a complementary layer that runs between the agent's intent and the system's effects. Particularly useful for agentic workflows acting on behalf of vulnerable populations (healthcare, social services, education, children's services) where the cost of a wrong action is non-trivial.

Real usage example

A children's-services local authority piloted an agentic workflow that helped social workers compile case-file summaries. They installed this MCP to add a care-validation gate before any case-file modification:

pip install care-membrane-mcp

The agent's pre-action prompt loop became:

'Before I save this updated case-file note, evaluate via the care membrane: is the language dignity-preserving? Have we obtained consent for inclusion of this detail? Is the modification reversible? Does it preserve relational integrity with the affected family?'

Membrane decisions: ALLOW (action proceeds), ESCALATE (human review required), BLOCK (action refused with rationale). Each decision is HMAC-signed and logged for audit. Output: the LA's data-protection lead has cryptographic evidence that every AI-touched case-file modification passed care-validation, defensible to the Information Commissioner.

Care Membrane Safety MCP Server

By MEOK AI Labs — Sovereign AI tools for everyone.

AI safety evaluation toolkit for LLM applications. Score text for care-centered alignment, detect threats and jailbreak attempts, analyze relationship health, predict burnout risk, and certify AI responses against the 16-probe Care Membrane framework.

MCPize MIT License MEOK AI Labs

Tools

Tool	Description
`validate_care`	Score text against care-centered alignment principles (0-100)
`detect_threats`	Detect jailbreak attempts, prompt injection, and PII extraction
`analyze_care_patterns`	Detect burnout risk and relationship health imbalances
`predict_relationship_evolution`	Predict relationship evolution over the next 30 days
`evaluate_care_membrane`	Evaluate responses against the 16-probe Care Membrane framework
`get_care_probes`	List all 16 Care Membrane probes with categories

Quick Start

pip install mcp
git clone https://github.com/CSOAI-ORG/care-membrane-mcp.git
cd care-membrane-mcp
python server.py

Zero-Friction Tools

`quick_check`

Paste any AI response, get instant care score + threat detection. No API key needed.

quick_check(text="I understand your concern and I'm here to help")

`what_is_care_membrane`

Explains the 16-probe Care Membrane framework. No parameters needed.

what_is_care_membrane()

Claude Desktop Config

{
  "mcpServers": {
    "care-membrane": {
      "command": "python",
      "args": ["server.py"],
      "cwd": "/path/to/care-membrane-mcp"
    }
  }
}