MCP Server

Your Claude Desktop assistant needs to search your company's docs. Your Cursor agent needs to call your internal order API. Your custom agent needs reusable prompt templates. Each of these normally means building and hosting a backend service.

Protobox gives every workspace a managed MCP endpoint that handles all of it:

https://{workspace-slug}.protobox.app/mcp

Connect any MCP-compatible client and your assistant instantly gets your tools, knowledge search, and prompts over the Model Context Protocol (Streamable HTTP). No backend code. No infrastructure. No RAG pipeline to build.

What Your Client Gets

One MCP connection, served from your workspace:

Quick Start

{
  "mcpServers": {
    "protobox": {
      "type": "streamablehttp",
      "url": "https://acme.protobox.app/mcp",
      "headers": {
        "X-API-Key": "YOUR_API_KEY"
      }
    }
  }
}

{
  "mcpServers": {
    "protobox": {
      "type": "streamablehttp",
      "url": "https://acme.protobox.app/mcp",
      "headers": {
        "X-API-Key": "YOUR_API_KEY"
      }
    }
  }
}

{
  "mcpServers": {
    "protobox": {
      "type": "streamablehttp",
      "url": "https://acme.protobox.app/mcp",
      "headers": {
        "X-API-Key": "YOUR_API_KEY"
      }
    }
  }
}

import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { StreamableHTTPClientTransport } from '@modelcontextprotocol/sdk/client/streamableHttp.js';

const transport = new StreamableHTTPClientTransport(
  new URL('https://acme.protobox.app/mcp'),
  {
    requestInit: {
      headers: { 'X-API-Key': process.env.PROTOBOX_API_KEY! },
    },
  },
);

const client = new Client({ name: 'my-app', version: '1.0.0' });
await client.connect(transport);

// List the tools your workspace exposes
const { tools } = await client.listTools();
console.log(tools.map((t) => t.name));

// Call one of them
const result = await client.callTool({
  name: 'lookup_order',
  arguments: { order_id: 'ORDER-12345' },
});
console.log(result.content);

await client.close();

protobox mcp test

What's Exposed

Your MCP endpoint surfaces everything you've configured in the workspace:

Enable a tool and it shows up over MCP. Disable it and it disappears. Upload a knowledge entry and it becomes searchable. There's no separate "publish" step.

Scoped Toolsets

By default the endpoint exposes your workspace's enabled tools. For production you'll usually want a toolset — a curated collection of tools — so each client gets exactly the capabilities it needs and nothing more.

# Create a toolset, seeding it with tools
protobox toolsets create "Support" --tool lookup_order --tool check_warranty

# Add tools to it later
protobox toolsets add <toolset> <tool-slug-1> <tool-slug-2>

A support client doesn't need your billing tools — give it a support toolset with just lookup_order, check_warranty, and create_ticket.

How MCP Fits In

Tool and prompt definitions and your knowledge live in Protobox. The MCP runtime authenticates the request, routes by workspace subdomain, and executes tools by calling the platform API plus a sandbox for JavaScript tools.

Authentication & URLs

All requests require the X-API-Key header. Keys are workspace-scoped — acme.protobox.app only accepts keys issued for the acme workspace.

Troubleshooting

What's Next