What is MCP?

The Model Context Protocol (MCP) is an open standard for connecting AI assistants to data sources and tools. The MyMemory MCP server gives your AI assistant tools to search, save, fetch and delete memories in your MyMemory library. There are two ways to connect, and you can mix both across different clients: Both expose the same five tools and the same MyMemory library. The local server is published on npm as @mymemory.dev/mcp (MyMemory is becoming Everstash).
Your API key gives the MCP connection access to your account. Never commit it, paste it into chat, include it in screenshots, or share a config containing it. Rotate any exposed key.

Quick start

1

Create an API key

Sign in at mymemory.dev and generate a key from the API key card on your dashboard. It starts with sm_.
2

Connect a client

Run the setup wizard below, click one of the one-click install buttons, or copy a config from Add memory → MCP in the app and paste it in yourself.
3

Restart your client

Restart the client or reload its MCP servers. The local server’s first start downloads the package, which takes a few seconds.

Set up in one command

The fastest way to connect any number of clients at once:
This finds which of Claude Desktop, Claude Code, Cursor, Windsurf, VS Code and Codex CLI are installed on your computer, asks which ones to connect and for your API key, then writes each one’s config itself — backing up whatever was already there. It needs an interactive terminal, since it asks you questions; restart the apps it touches when it’s done.

One-click install

Cursor and VS Code can install an MCP server directly from a link, no terminal or manual JSON needed. Open Add memory → MCP in the app and click Add to Cursor or Add to VS Code; the app fills in the server for you, and you paste in your API key once inside the client.

Remote MCP server (no install)

For clients that can add a remote MCP server by URL — with no local Node.js, npx, or config file — point them at:
Authenticate with an Authorization: Bearer sm_your_api_key_here header, same as the REST API. This is a stateless Streamable HTTP endpoint: every request carries its own auth, and there’s no server-side session to keep alive. It’s the option to reach for on Claude.ai connectors, ChatGPT connectors, or any other client that doesn’t run a local process.

Configuration by client

Claude Desktop, Cursor, Windsurf, Cline and most others

Claude Code

--scope user registers the server for every project. Omit it and Claude Code only wires it into whatever project directory the command runs in.

Codex CLI

VS Code

Save this as .vscode/mcp.json. VS Code asks for your API key the first time and stores it securely, so the key never sits in the file:
Or click Add to VS Code from Add memory → MCP in the app for the one-click deep link instead.

Tool reference

Both the local and remote server expose the same five tools:
Confirmed deletion is irreversible. Review the preview returned by the first delete_memory call before sending confirm: true.

Environment variables (local server)

Older configs that run a local build with MYMEMORY_API_KEY keep working; switching to npx -y @mymemory.dev/mcp gets you updates automatically.

Troubleshooting

  • No API key configured: add EVERSTASH_API_KEY to the env block of your config (local) or the Authorization: Bearer header (remote).
  • 401 or Invalid API key format: replace the placeholder with a valid sm_... key.
  • npx not found, or the server never starts: install Node.js 18+ and restart the client. On Windows, if your client still can’t start npx, use "command": "cmd" with "args": ["/c", "npx", "-y", "@mymemory.dev/mcp"].
  • Tools don’t appear: fully restart the client and check its MCP logs.
  • Search is slow: large libraries can take a few seconds; allow the request to finish.
  • npx @mymemory.dev/mcp setup says it needs an interactive terminal: run it directly in a terminal window rather than through a script or CI job, or configure the client manually using the sections above.
Don’t put secrets in memory content, URLs or metadata. Tool errors can appear in your client’s local logs.