一个受治理的 Model Context Protocol 服务器，通过单一 MCP 契约向两个独立的 AI 聊天应用提供课程、批次、费用、潜在学员和回调能力。
一个 MCP 服务器、两个独立客户端（学员 + 顾问）——无需重复集成
数据库凭据、认证、审计、验证——全部集中在服务器端
写入操作需要确认门禁（prepare → confirm）——LLM 无法单独创建潜在学员
每次工具调用都会被审计（操作者、客户端、参数哈希、结果、延迟）
一个无 MCP 对比演示展示了没有 MCP 时你会失去什么
Related MCP server: ComplyOS
# 1. PostgreSQL
docker compose up -d postgres
# 2. Migrate + seed
uv sync
uv run alembic upgrade head
uv run python scripts/seed_demo.py
# 3. Start services (4 terminals)
uv run uvicorn services.mcp_server.app:asgi_app --port 8010
uv run uvicorn services.learner_host.api:app --port 8020
uv run uvicorn services.counsellor_host.api:app --port 8030
uv run streamlit run ui/app.py --server.port 8501打开 http://localhost:8501——两个聊天标签页（学员 + 顾问）。
无 MCP 对比：
uv run streamlit run ui/no_mcp_demo.py --server.port 8502
提示词 | 结果 |
| 列出 4 门课程 |
| 批次日期 + 费用报价 + 政策 |
| 返回政策文本 |
| ✅/❌ 创建潜在学员前的确认门禁 |
提示词 | 结果 |
| 列出 4 门课程 |
| 3 个有名额的批次 |
| 报价 ID + 总计（INR） |
| 显示已分配的潜在学员 |
| ✅/❌ 确认门禁 |
端口 | 服务 | 角色 |
5433 | PostgreSQL | 权威数据源（课程、批次、潜在学员、审计） |
8010 | MCP Server | 工具（11）+ 资源（8）+ 提示词（2）、JWT 认证、RBAC、审计 |
8020 | Learner Host | LangGraph 应用——学员 JWT、写入确认门禁 |
8030 | Counsellor Host | LangGraph 应用——顾问 JWT、潜在学员管理 |
8501 | Streamlit UI | 两个聊天标签页（基于 MCP） |
8502 | No-MCP Demo | 相同流程、直连数据库——展示 MCP 所防护的风险 |
技术栈： Python 3.11 · MCP SDK · LangGraph · FastAPI · SQLAlchemy 2 · PostgreSQL 16 · Pydantic v2 · Ollama (qwen3.5:2b) · Streamlit
概念 | 位置 | 重要性 |
确认门禁 |
| 没有人工 ✅，LLM 无法创建潜在学员 |
幂等性 |
| 网络重试不会产生重复数据 |
RBAC |
| 学员无法查看他人的潜在学员 |
审计 |
| 每次调用均记录：谁、什么操作、结果、延迟 |
无状态性 | 服务器生成的 ID（ | 无需会话即可水平扩展 |
scai-mcp-admissions/
├── services/
│ ├── mcp_server/ # MCP server (tools, resources, prompts, auth, audit)
│ ├── learner_host/ # LangGraph learner app (port 8020)
│ └── counsellor_host/ # LangGraph counsellor app (port 8030)
├── ui/
│ ├── app.py # Streamlit — 2 chat tabs (MCP)
│ └── no_mcp_demo.py # Streamlit — no-MCP comparison (direct DB)
├── packages/
│ ├── contracts/ # Pydantic tool inputs/outputs + domain models
│ ├── shared/ # Config, LLM adapter, JWT tokens
│ └── observability/ # Structured logging, tracing
├── scripts/
│ ├── seed_demo.py # Seed 4 courses, 4 batches, 4 fee plans, 3 policies
│ ├── issue_dev_token.py # Issue dev JWTs for manual testing
│ └── run_demo_checks.py # Smoke tests against running server
├── tests/ # unit, contract, integration, security, e2e
├── migrations/ # Alembic migrations
├── data/demo_seed/ # Seed data + knowledge_base.json
├── mcp_concept.ipynb # MCP concept notebook (what/why/how/scale/use cases)
├── mcp_flow_diagram.mmd # Mermaid source for architecture diagram
├── mcp_flow_diagram.png # Rendered architecture diagram
├── RUN_GUIDE.md # Step-by-step run guide with test prompts
└── docker-compose.yml # PostgreSQL 16uv run pytest # all tests
uv run pytest -m unit # just unit tests
uv run pytest -m contract # contract testsmcp_concept.ipynb——完整的 MCP 概念讲解
This server cannot be installed
Unclaimed servers have limited discoverability.
If you are the server author, to access and configure the admin panel.
Runtime AI governance: decision gates, human approval, hash-chained audit, compliance mapping.
Enterprise AI Control Plane: governance, guardrails, spend tracking, compliance & smart routing.
Odoo ERP for AI agents: hosted OAuth endpoint, gated writes, one endpoint for every instance.
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/avivek-dwivedi/custom-mcp-server'
If you have feedback or need assistance with the MCP directory API, please join our Discord server
