Set Up an MCP Memory Server in 10 Minutes

Real config, for Claude Code, Claude Desktop, Cursor, and Windsurf — plus the self-hosted path.

By Anthony Conti · Astra AI, LLC · Last updated September 7, 2026

Setting up an MCP memory server means adding one entry to your AI tool's MCP configuration — a URL for the remote path, or a local command for the self-hosted path — after which every tool call your agent makes can read and write persistent memory. Below is the exact, verified setup for MIND, an MCP memory server (see what an MCP memory server is if you landed here first), across the four most common MCP hosts. Total time for the remote path: under 2 minutes per tool. The self-hosted path takes closer to 10, mostly waiting on npm install.

What do you need before you start?

An account — free tier, no credit card, at m-i-n-d.ai.

One of: Claude Code, Claude Desktop, claude.ai, Cursor, or Windsurf already installed.

For the self-hosted path only: Node.js 18 or higher (node -v to check).

You do not need to decide between the remote and self-hosted paths before starting — each of the four tool sections below shows exactly one path, and the self-hosted alternative is broken out separately further down for anyone who needs it. Most readers should just follow the section for whichever tool they already have open.

Claude Code — one command

This is the fastest of the four. Paste the command, then run /mcp and click through the browser sign-in.

claude mcp add --transport http mind https://www.m-i-n-d.ai/mcp

Then run /mcp inside Claude Code. Your browser opens to MIND — sign in if needed, click Connect, and you're done. Claude Code now has full read/write access to your MIND: memory, tasks, contacts, research.

Claude Desktop & claude.ai

Go to Settings → Connectors → Add custom connector and enter https://www.m-i-n-d.ai/mcp as the URL. Click through sign-in, hit Connect, and MIND is available in every conversation from then on.

One-click alternative: download the mind.mcpb bundle from m-i-n-d.ai/downloads/mind.mcpb and double-click it — Claude Desktop opens and installs it, then prompts you for a MIND API key.

Cursor

Add MIND to .cursor/mcp.json in your project (or the global config) with the same URL used above:

{
  "mcpServers": {
    "mind": {
      "url": "https://www.m-i-n-d.ai/mcp"
    }
  }
}

Reload the window. On first tool call, your browser opens for the same one-click Connect flow described above.

Windsurf

Same JSON shape, different file: ~/.codeium/windsurf/mcp_config.json.

{
  "mcpServers": {
    "mind": {
      "url": "https://www.m-i-n-d.ai/mcp"
    }
  }
}

Restart Windsurf (or reload the MCP panel) and connect the same way.

Self-hosted / local install (for offline, air-gapped, or key-based setups)

Prefer to run the server locally over stdio and manage a key by hand? This is the open-source npm path (@astramindapp/mcp-server, currently at version 0.23.0), and it's the one to use for self-hosted MIND deployments or air-gapped machines.

Get a key. Sign up free at m-i-n-d.ai, then go to Settings → Developer → Create API Key. It starts with mind_.
Install (Node 18+ required).
```
npm install -g @astramindapp/mcp-server
```
Wire it in. For Claude Code:
```
claude mcp add mind -- env MIND_API_KEY=mind_your_key_here mind-mcp
```
Or, for any MCP-compatible tool that reads a JSON config (Claude Desktop, Cursor, Windsurf, VS Code):
```
{
  "mcpServers": {
    "mind": {
      "command": "mind-mcp",
      "env": {
        "MIND_API_KEY": "mind_your_key_here",
        "MIND_BASE_URL": "https://m-i-n-d.ai"
      }
    }
  }
}
```
For Claude Desktop this JSON goes in claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json; Windows: %APPDATA%\Claude\claude_desktop_config.json). Fully quit and reopen the app afterward — it only reads this file on launch.

Remote or local — which should you pick?

Both paths connect to the exact same MIND account and the exact same memory — the difference is entirely about where the server process runs and how you authenticate:

Pick remote (the one-paste path) if you want the fastest possible setup, you are fine with the server running on MIND's own infrastructure, and you want the key rotated and revoked for you automatically through browser OAuth.

Pick local/self-hosted (the npm path) if you are on an air-gapped machine, you are running a self-hosted MIND deployment and need to point at your own MIND_BASE_URL, or your organization requires you to manage API keys explicitly rather than through OAuth.

Either way, you are running the same open-source code. The published npm package (@astramindapp/mcp-server, MIT licensed) is what both paths run underneath — remote just runs it for you.

How do you know it worked?

Ask your agent to call a MIND tool directly — for example, "call mind_context and tell me what you get back." A working connection returns real content: identity, rules, recent activity. A broken one returns a tool-not-found error (the connector never registered) or an auth error (the key or OAuth session is invalid) — both point you back to the step above for your specific tool.

Troubleshooting

"Connect" opened the website instead of connecting. Use the www host exactly: https://www.m-i-n-d.ai/mcp. A non-www variant of this endpoint had a bug, fixed in July 2026.

My tool has no "Connectors" UI. Not every MCP host ships one yet. Use the Claude Code one-liner if you're on Claude Code, or the local npm install below — it works with any MCP-compatible tool via a config file.

npx can't find the package, or it hangs. Install globally first (npm install -g @astramindapp/mcp-server), then reference the local binary — or run the one-shot setup wizard: npx -y @astramindapp/mcp-server@latest mind-mcp-setup.

Frequently asked

Do I need an API key to connect?

Not for the one-paste remote flow. Clicking Connect in your browser mints a scoped access key automatically, the same kind of key you would otherwise create by hand. You only need to create a mind_-prefixed key yourself for the local stdio/npm path.

How do I know it actually connected?

Ask your agent something only your MIND would know — "what do you know about me?" or "call mind_context." A working connection returns real content back; a broken one returns a tool-not-found or auth error, which is your signal to redo the step above.

Can I connect more than one MIND account?

Yes. Add the connector once per account, giving each a distinct name (mind-personal, mind-work). Each Connect click opens sign-in where you choose which account to authorize.

Does this work with a self-hosted MIND instance?

Yes — the local npm path takes an optional MIND_BASE_URL environment variable that overrides the default https://m-i-n-d.ai host, for exactly this case.

The self-hosted path installs the open-source @astramindapp/mcp-server package on npm (source on GitHub), which requires Node.js 18 or higher — get it from nodejs.org if node -v comes back below that.

Try MIND free →What is an MCP memory server? →

One memory. Every AI.

MIND is an MCP memory server backed by a real knowledge graph — the memory layer Claude, GPT, Gemini, Cursor and Windsurf all read from. Free tier, no credit card.

Try MIND free →

© 2026 Astra AI, LLC · m-i-n-d.ai
MCP memory server · Setup guide · Agent memory · Persistent memory · Second brain · Comparisons · MCP docs