MCP Server

Planned Feature

The MCP Server is planned for a future release. This documentation describes the intended functionality to help you plan your integration ahead of time.

What is MCP?

The Model Context Protocol (MCP) is a standard for AI tools to share context. Cortex will provide an MCP server that enables tools like Cursor, Claude Desktop, and custom apps to share memory.

Result: Tell one tool something, all tools remember it!

---

Compatible Tools

Cursor

Claude Desktop

Windsurf

Custom Apps

---

Planned Setup

1

Install Cortex CLI

Terminal

$ brew install cortex-memory/tap/cli

Terminal

$ npm install -g @cortexmemory/cli

2

Start MCP server

Terminal

$ cortex mcp start

3

Configure Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "cortex": {
      "command": "cortex",
      "args": ["mcp", "serve"]
    }
  }
}

4

Restart Cursor

Restart Cursor to load the MCP server.

1

Install Cortex CLI

Terminal

$ brew install cortex-memory/tap/cli

Terminal

$ npm install -g @cortexmemory/cli

2

Configure Claude Desktop

Add to Claude Desktop config:

{
  "mcpServers": {
    "cortex": {
      "command": "cortex",
      "args": ["mcp", "serve"]
    }
  }
}

Config locations: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), %APPDATA%\Claude\claude_desktop_config.json (Windows), ~/.config/Claude/claude_desktop_config.json (Linux)

3

Restart Claude Desktop

Restart Claude to load the MCP server.

import axios from 'axios';

const MCP_ENDPOINT = 'http://localhost:3000';
const USER_ID = 'user-123';

// Store memory
async function remember(content: string) {
  await axios.post(`${MCP_ENDPOINT}/add_memories`, {
    memory: content,
    user_id: USER_ID,
  });
}

// Retrieve context
async function search(query: string) {
  const response = await axios.post(`${MCP_ENDPOINT}/search_memory`, {
    query,
    user_id: USER_ID,
    limit: 5,
  });
  return response.data.results;
}

---

How MCP Memory Works

Hive Mode in Action

When using MCP, all your AI tools share ONE memory space (Hive Mode). Cursor stores a preference, Claude can read it!

MCP Memory Architecture

AI Tools

Cortex MCP Server

Shared Memory Space

Example Flow

1. In Cursor: You say "I prefer TypeScript for backend"
2. Cursor → MCP: Stores preference via add_memories
3. Later, in Claude: You ask about coding advice
4. Claude → MCP: Searches for "coding preferences"
5. Claude: Uses that fact to personalize response

Result: Claude knows your preferences without you repeating them!

---

Planned MCP Endpoints

Endpoint	Method	Description
cortex_remember / /add_memories	POST	Store a memory
cortex_search / /search_memory	POST	Search memories
cortex_recall / /list_memories	GET	Get context for LLM
cortex_get_facts / /get_memory	GET	Retrieve extracted facts
/delete_all_memories	POST	Clear user memories

Store Memory

POST /add_memories
{
  "memory": "User prefers dark mode",
  "user_id": "user-123",
  "metadata": { "source": "cursor", "importance": 70 }
}

Search Memory

POST /search_memory
{
  "query": "user preferences",
  "user_id": "user-123",
  "limit": 5
}

---

Deployment Options

Run MCP server locally on your machine:

Terminal

$ cortex mcp start

Tip

Free, data stays local, requires server running.

Environment Variables:

CONVEX_URL=https://your-project.convex.cloud
MCP_PORT=3000
MCP_LOG_LEVEL=info

Cloud-hosted MCP endpoint—no local server needed:

{
  "mcpServers": {
    "cortex-cloud": {
      "url": "https://mcp.cortex.cloud/v1",
      "auth": {
        "type": "bearer",
        "token": "cortex_sk_your_api_key"
      }
    }
  }
}

Premium Benefits

• Always available (99.9% uptime)
• Auto-embeddings (no API key needed)
• Auto-fact extraction
• Team sharing
• Analytics dashboard

---

User Isolation

Automatic Isolation

Each user_id has completely isolated memories. User A cannot see User B's data.

// User A stores preference
POST /add_memories
{ "memory": "I prefer light mode", "user_id": "alice@example.com" }

// User B stores preference
POST /add_memories
{ "memory": "I prefer dark mode", "user_id": "bob@example.com" }

// Searches are isolated
POST /search_memory
{ "query": "theme preference", "user_id": "alice@example.com" }
// Returns: "I prefer light mode" (Alice's only)

---

Use Cases

Cross-IDE Preferences

1. In Cursor: "I prefer 2-space indentation"
   → Stores in MCP

2. Later, in Windsurf: Ask AI to generate code
   → Queries MCP, gets "2-space indentation" preference
   → Generates code with correct indentation

3. In Claude: "Help me write a config file"
   → Knows your preferences from MCP

Personal Knowledge Base

Day 1 (Cursor): "Working on Project Apollo, a React app"
Day 3 (Claude): "Apollo uses PostgreSQL"  
Day 7 (Any tool): "Tell me about Apollo"
→ MCP returns all accumulated facts
→ All tools have complete context

---

Related Features