MCP Overview

What is MCP?

The Model Context Protocol (MCP) is an open standard for connecting AI assistants to external tools and data sources. Remb implements an MCP server that exposes 100 tools across 21 modules — memory, context, scanning, conversation tracking, plans, graph traversal, cross-project search, scratchpad, briefings, role contracts, skills, and more.

The key idea: your project's context is not a snapshot of its source files — it's a graph that your AI builds and curates over every session. Memories capture why decisions were made. Feature notes document what each part does. Conversation logs record what was accomplished. Scanning adds structural breadth — but it's optional and additive, not the foundation.

Setup

Via CLI (recommended)

The CLI includes a built-in MCP proxy that connects via stdio:

remb serve

This proxies JSON-RPC messages between stdio and the Remb HTTP API. No tool definitions are hardcoded — the CLI dynamically forwards everything the server supports.

Direct HTTP

You can also connect directly to the HTTP MCP endpoint:

POST https://www.useremb.com/api/mcp
Authorization: Bearer rmb_your_api_key
Content-Type: application/json

{ "jsonrpc": "2.0", "method": "tools/list", "id": 1 }

IDE Configuration

VS Code (GitHub Copilot)

{
  "servers": {
    "remb": {
      "command": "remb",
      "args": ["serve"],
      "env": {}
    }
  }
}

Claude Code

{
  "mcpServers": {
    "remb": {
      "command": "remb",
      "args": ["serve"]
    }
  }
}

Cursor

{
  "mcpServers": {
    "remb": {
      "command": "remb",
      "args": ["serve"]
    }
  }
}

How It Works

1. Your IDE starts the MCP server via remb serve
2. The AI assistant discovers available tools via tools/list
3. When the AI needs context, it calls tools like remb__memory_load_context
4. The CLI proxies the request to the Remb API and returns the result
5. All data is persisted server-side across sessions, IDEs, and machines

Tool Budget

100 tools is too many to keep loaded at once for a focused task. Pass _meta.tool_budget on tools/list to get the top-K most relevant tools for the current task. Optionally pass _meta.context_hint with a short description of what you’re working on to score relevance. Always-included tools (session_start, memory_load_context, tool_discover) are guaranteed even under tight budgets.

{
  "method": "tools/list",
  "params": {
    "_meta": {
      "tool_budget": 20,
      "context_hint": "refactor authentication"
    }
  }
}

Tool Categories