nexus-mcp
MCP server that connects to a sales CRM database, enabling AI agents to search prospects, view pipeline status, manage follow-ups, and update records using natural language commands.
README
nexus-mcp
An MCP (Model Context Protocol) server that exposes a real sales CRM as tools an AI agent can call — plus a CLI that does the same job with no model involved.
Built to solve an actual problem, not as an exercise. I run a small web-development studio and my CRM had grown to ~270 prospects with long free-text notes that nobody re-read. Prospects who had shown interest were quietly dying in the list, because a spreadsheet only shows you what you remember to look for. The first time I ran this, it surfaced two leads I had abandoned a week earlier.
$ nexus-crm hoy
VENCIDOS (1)
────────────
#a3f21c08-4d19-4b7e-9e02-6c1f8b40d772 Estudio Ferrari & Asoc. · Legal · Zárate [Respondió]
contacto: contacto@ferrari.example.com · seguir: 2026-07-24 · $450000
vencido hace 3 día(s)
ABANDONADOS — nadie les puso fecha (4)
──────────────────────────────────────
#7b90e4d1-2c53-49aa-8f61-0d5e7a3c9b14 Odontología Norte · Salud · Zárate [Respondió]
contacto: +54 9 3487 000002 · $590000
...
The output above comes from
demo/seed.sql— invented businesses, invented phone numbers, reservedexample.comdomains. No client data lives in this repository.
The five tools
| Tool | What it does |
|---|---|
buscar_prospectos |
Free-text search across name, industry, area, contact and notes, with optional filters by status/industry/area |
resumen_pipeline |
Pipeline snapshot: count per status, overdue follow-ups, abandoned leads, total value at stake |
seguimientos_pendientes |
Who to chase: overdue, upcoming, and abandoned — interested, not discarded, and with no follow-up date set |
cambiar_estado |
Move a prospect along the pipeline |
agregar_nota |
Append a dated entry to the prospect's history, in the same format the CRM's own UI writes |
The third one is the reason this exists. Overdue follow-ups are easy — any list view shows them. Leads with no date at all are invisible by construction, and those are the ones that cost money.
Stack
TypeScript · Node 22 · @modelcontextprotocol/sdk · Supabase (PostgreSQL) · Zod
Tool inputs are validated with Zod schemas, so the model gets a typed contract and a useful error instead of a silent failure. The server speaks MCP over stdio, which is what desktop MCP clients expect.
Row-level security
The prospectos table has RLS enabled. Worth knowing, because it fails quietly: with the publishable key alone, Supabase returns an empty list rather than an error — so the tools look like they work and report zero results.
The server signs in with email/password before its first query and holds an authenticated session, which is what the RLS policy checks. Credentials live in .env, never in the repo.
Running it
npm install
cp .env.example .env # fill in your Supabase project and login
npm run build
As a CLI:
node dist/cli.js hoy
node dist/cli.js resumen
node dist/cli.js buscar odonto
node dist/cli.js estado 42 "Respondió"
node dist/cli.js nota 42 "resent the quote, he'll look at it this week"
As an MCP server, add it to your client's config:
{
"mcpServers": {
"nexus-crm": {
"command": "node",
"args": ["/absolute/path/to/nexus-mcp/dist/server.js"]
}
}
}
Then ask in plain language: "who should I follow up with today?", "mark Ferrari as replied and note that he asked for the quote again".
Trying it without my data
demo/seed.sql creates the table, its RLS policy and 15 fictional prospects — invented names, phone numbers and negotiation notes. Run it in an empty Supabase project, point your .env at it, and everything below works against that instead.
What this is not
It doesn't replace the CRM and it doesn't collect data on its own. It's a query-and-update layer on top of a database that already existed. The CRM UI stays the place where records are created; this is for the question a UI is bad at answering: what am I forgetting?
License
MIT
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。