VoxPilot MCP Server
A production-ready remote MCP server for the VoxPilot AI SaaS platform, providing 15 tool groups for restaurant orders, messaging, POS, calendars, customer memory, QA, and handoff.
README
VoxPilot MCP Server
Production-ready remote Model Context Protocol (MCP) server for the VoxPilot AI SaaS platform.
| Transport | Streamable HTTP (remote MCP, not stdio) |
| Auth | JWT (RS256), tenant-scoped, scope-based permissions |
| Modules | 15 tool groups covering restaurant orders, messaging, POS, calendars, customer memory, QA, handoff, and more |
| Deployment | Docker → GHCR → Coolify |
| Runtime | Node.js 22, Express 5, TypeScript |
Quick Start
# Local dev
npm install
cp .env.local.example .env
npm run dev
# Docker
npm run docker:build
npm run docker:run
Endpoints
| Method | Path | Auth | Description |
|---|---|---|---|
| POST | /mcp |
JWT | MCP protocol endpoint (Streamable HTTP) |
| GET | /health |
No | Health check |
| GET | /ready |
No | Readiness check |
| GET | /live |
No | Liveness probe |
| GET | /version |
No | Server version info |
| GET | /tools |
No | Module diagnostics (not MCP protocol) |
| GET | /tools/:module |
No | Per-module diagnostics |
| GET | /metrics |
No | Prometheus metrics (if enabled) |
| GET | /.well-known/oauth-protected-resource |
No | OAuth metadata |
MCP Endpoint
The real MCP protocol happens at /mcp. The /tools/:module URLs are module aliases for diagnostics, documentation, and backend configuration clarity — they do not replace /mcp.
Local Caddy
mcp.voxpilot.test {
tls internal
reverse_proxy mcp-server:8080
}
Add to /etc/hosts:
127.0.0.1 mcp.voxpilot.test
Auth Model
JWT tokens issued by VoxPilot NestJS backend with these claims:
{
"iss": "voxpilot-api",
"aud": "voxpilot-mcp",
"sub": "user_or_service_id",
"tenantId": "tenant_uuid",
"role": "owner | admin | agent | system",
"scopes": ["mcp:tools:read", "orders:read", "..."],
"callId": "optional",
"sessionId": "optional"
}
Dev Bypass
Set DEV_BYPASS_AUTH=true in .env for local development (blocked in production).
DEV_BYPASS_AUTH=true
DEV_TENANT_ID=dev-tenant
DEV_SUBJECT=dev-user
DEV_ROLE=system
DEV_SCOPES=mcp:tools:read,orders:read,orders:write,messages:send,pos:read,pos:write,customers:read,customers:write,calendar:read,calendar:write,business:read,payments:write,qa:write,handoff:write,analytics:read
Tool Modules (15)
| # | Module | Status |
|---|---|---|
| 1 | system-core |
3 tools implemented (ping, version, listModules) |
| 2 | restaurant-order |
1 mock tool (validateOrderDraft) |
| 3 | messaging-dispatch |
1 mock tool (dispatchTest) |
| 4 | pos-connector |
1 mock tool (testConnection) |
| 5 | customer-memory |
Placeholder |
| 6 | calendar-appointments |
1 mock tool (getProviderStatus) |
| 7 | business-knowledge |
Placeholder |
| 8 | payments-deposits |
Placeholder |
| 9 | call-qa |
Placeholder |
| 10 | handoff |
Placeholder |
| 11 | delivery-routing |
Placeholder |
| 12 | promotions-upsell |
Placeholder |
| 13 | menu-availability |
Placeholder |
| 14 | onboarding |
Placeholder |
| 15 | analytics-insights |
Placeholder |
Integration
VoxPilot NestJS backend calls the MCP server via POST to /mcp with:
Authorization: Bearer <jwt>
Content-Type: application/json
Accept: application/json, text/event-stream
Example JSON-RPC tool call:
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "systemCore.ping",
"arguments": { "message": "hello" }
}
}
Spec-Driven Development
Each module is implemented spec-by-spec. See specs/ for the 16 spec files (000–015).
No module is implemented without a spec.
What is intentionally not yet implemented
- Real WhatsApp, Telegram, Slack, SMS, email dispatch
- Real POS connections (TastyIgniter, Odoo)
- Real Google Calendar / Outlook integration
- Real payment processing (Stripe)
- Real call QA analysis
- Real customer memory persistence
- Real analytics queries
- Redis-backed queues or caching
- PostgreSQL audit persistence
- OpenTelemetry export
Production Deployment
# docker-compose.yml (VoxPilot root stack)
mcp-server:
build:
context: ./voxpilot-mcp-server
dockerfile: docker/Dockerfile
ports:
- "8080:8080"
env_file:
- .env
networks:
- voxpilot
Image published to ghcr.io/nehilor/voxpilot-mcp-server with tags latest, develop, sha-<shortsha>.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。