emotional-deescalation-mcp 0.1.0

Universal MCP server for emotional tone analysis and de-escalation in communications

Emotional De-escalation MCP Server v2

Universal MCP server for emotional tone analysis and de-escalation using a 5-axis communication style model.

Based on: Mistakes to Avoid When Developing Chatbots for User Support

5-Axis Communication Style Model

Every message is characterized by a style vector — 5 independent axes on a discrete scale from -2 to +2:

Axis	-2	-1	0	+1	+2
W Warmth	cold, detached	cool	neutral	friendly	warm, empathetic
F Formality	slang, crude	casual	balanced	professional	formal, official
P Playfulness	dead serious	dry	balanced	light, witty	humorous, ironic
A Assertiveness	uncertain, meek	tentative	balanced	confident	demanding, forceful
E Expressiveness	terse, reserved	restrained	balanced	animated	emotional, intense

Style Combinations

Style	W	F	P	A	E
Sarcasm	-2	-1	+2	0	+1
Friendly humor	+1	-1	+2	-1	0
Flirtatious	+2	-2	+1	-1	+1
Business tone	0	+2	-2	0	-1
Aggression	-2	-2	-2	+2	+2
Desperation	0	-1	-2	-1	+2
Passive-aggression	-1	0	+1	+1	-1

De-escalation Strategy

Instead of a single coefficient, de-escalation operates per-axis:

Axis	Shift	Rationale
Warmth	+1	Increase empathy
Formality	+1	Slightly more professional
Playfulness	→ 0	Reduce sarcasm risk
Assertiveness	-1	Reduce pressure
Expressiveness	-1	Calm down intensity

This breaks the positive feedback loop by ensuring the bot's response vector is always shifted toward a more constructive zone.

Tools

emotion_analyze

Analyze a message → emotion + style vector + style label.

{
  "text": "ну конечно, ваш замечательный бот мне так помог, спасибо огромное",
  "language_hint": "ru"
}

Returns:

{
  "emotion": "anger",
  "intensity": 1,
  "style_vector": {"warmth": -2, "formality": -1, "playfulness": 2, "assertiveness": 0, "expressiveness": 1},
  "detected_style": "sarcasm",
  "explanation": "...",
  "triggers": ["ну конечно", "замечательный", "спасибо огромное"]
}

emotion_de_escalate

Rewrite a draft to match a target style vector.

Auto mode (default): analyzes user, applies de-escalation shifts. Override mode: pass target_style explicitly.

{
  "user_message": "what the hell is wrong with the delivery?!",
  "draft_response": "Your order should arrive by April 10.",
  "target_style": {"warmth": 1, "formality": 1, "playfulness": 0, "assertiveness": 0, "expressiveness": 0}
}

emotion_evaluate_dialogue

Evaluate full dialogue → per-message vectors + trend + feedback loop risk.

{
  "messages": [
    {"role": "user", "text": "hello"},
    {"role": "bot", "text": "Hello! How can I help?"},
    {"role": "user", "text": "what the hell is wrong with the delivery?!"},
    {"role": "bot", "text": "Your order #3756 arrives April 10."},
    {"role": "user", "text": "ok I guess I'll wait"}
  ],
  "response_format": "markdown"
}

Returns a table:

#	Role	Emotion	W	F	P	A	E	Style
1	user	neutral	0	-1	0	0	0	casual
2	bot	neutral	+1	0	0	0	0	friendly
3	user	anger	-2	-2	-2	+2	+2	aggressive
4	bot	neutral	0	+1	-1	0	-1	business
5	user	neutral	0	-1	0	-1	0	resigned

Setup

pip install -e .
export ANTHROPIC_API_KEY=sk-ant-...
python server.py

Claude Desktop / Claude Code config

{
  "mcpServers": {
    "emotional-deescalation": {
      "command": "python",
      "args": ["/path/to/emotional-mcp/server.py"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Agent integration

For LLM agents connecting via MCP: see AGENTS.md — a compact instruction file designed to minimize context window usage while providing all necessary tool selection and invocation guidance.

Architecture

MCP Client (Claude Desktop, Claude Code, any MCP host)
    │ stdio
    ▼
emotional_deescalation_mcp
    ├── emotion_analyze        → emotion + style vector W/F/P/A/E
    ├── emotion_de_escalate    → rewrite draft to target vector
    └── emotion_evaluate_dialogue → per-message vectors + dynamics
            │
            ▼
      Anthropic Claude API (analysis backend)

License

MIT