social-skills-coach-mcp

中文說明請見 README-cht.md。

An MCP server that turns a PEERS-style social-skills curriculum into a full coaching loop — Analyze → Coach → Role-Play → Reflect — that any MCP client can run with its own model.

It surpassed 1,600 downloads in its first week, with no promotion — suggesting genuine demand for accessible social-skills tooling.

It is the standalone, distributable form of the Social Skills AI Coach project (try the live demo or video): describe a situation, get concrete curriculum-grounded advice, rehearse it in a role-play, and receive a structured reflection. The curriculum ships inside the package, synced at build time from the project's single-source Agent Skill (skills/social-skills-coach/).

Why bring your own model

The four agents are exposed as MCP prompts — they execute on the connecting client's model. The server only serves prompts + curriculum knowledge and runs no inference itself, so:

• No API key needed by this server.
• Works with any model your client supports (Anthropic, OpenAI, Google Gemini, local models, etc.) — plug in a more capable model than Demo Web App.
• No data stored. The server reads curriculum markdown and returns text; it persists nothing.

What it exposes

• Prompts (run on your model):
  • analyze_situation(situation) — structure a social situation (who/what/where, channel, scenario type, goal) without giving advice yet
  • coach(situation) — concrete, curriculum-grounded advice and exact phrases you can say
  • roleplay(scenario) — plays the other person so you can practice; reacts to your skill level
  • reflect(transcript) — rubric-based, per-dimension evaluation of a role-play
• Tools (curriculum knowledge for grounding):
  • list_social_topics() — list the 12 curriculum topics (call this first)
  • get_social_knowledge({ topics }) — fetch the verbatim curriculum slice(s)

The agents reply in the same language the user writes in, so you can practice in your mother tongue. Language quality depends on the model you connect; tested working in English, Chinese, and Spanish on MiMo / DeepSeek.

Use with Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "social-skills-coach": {
      "command": "npx",
      "args": ["-y", "social-skills-coach-mcp"]
    }
  }
}

Other MCP clients (Cursor, Antigravity, …) use the same command + args.

Run directly

npx -y social-skills-coach-mcp        # speaks MCP over stdio

Or inspect it with the MCP Inspector:

npx @modelcontextprotocol/inspector npx -y social-skills-coach-mcp

Install globally (optional)

npm install -g social-skills-coach-mcp
# or: pnpm add -g social-skills-coach-mcp
social-skills-coach-mcp               # runs the stdio server

🔒 Security

• No secrets to leak. The server holds no API key and runs no inference — all reasoning happens on the connecting client's model. There is nothing here to exfiltrate.
• zod validation at the trust boundary. Every tool/prompt argument is validated with zod (topic keys are a fixed enum, min 1); malformed or out-of-range input is rejected before use.
• Stateless by design. It reads the bundled curriculum markdown and returns text — no database, no persistence, no telemetry.

⚠️ Disclaimer

This project is a conceptual product (minimum viable product) developed for the Kaggle AI Agents: Intensive Vibe Coding Capstone Project. The participating track is Agents for Good, and it is for review and research by interested parties only. All functions cannot replace professionally trained and licensed psychologists or therapists, and cannot provide any medical treatment or consultation.

The demo video and screenshots are in English for the Kaggle review. You can talk to the AI in your own language — how well it does depends on the model. I tested English, Chinese, and Spanish on MiMo / DeepSeek, and all worked smoothly.

Please always remember: you are talking to an AI. Avoid mentioning personal information such as your real name, phone number, or address; use a pseudonym if needed. AI can make mistakes and hallucinate — all suggestions are for reference only.

License

MIT