Stash / Blog · June 2026 · 7 min read

Claude MCP Connectors Explained: How to Add External Tools to Claude

Claude is a capable model, but out of the box it can't read your files, query your databases, or remember things between conversations. MCP connectors change that. This guide explains what they are, how to add one in about two minutes, and what to expect when you do.

What this covers: what MCP is and why it matters, the difference between local and remote connectors, how to add a custom MCP server to Claude.ai and Claude Desktop, and the easiest remote connector to try today.

What is MCP?

MCP stands for Model Context Protocol. It's an open standard (published by Anthropic) that defines how AI models talk to external tools and data sources. An MCP server exposes a set of named tools — functions the model can call — and the model decides when to call them based on what the user asks.

Before MCP, giving Claude access to external data meant either pasting it all into the context window (expensive, limited by the window size) or using proprietary API integrations (brittle, hard to build). MCP gives AI models a standardised way to call out to anything: a database, a web search tool, a calendar, a file system, a custom record store.

The connector is just a URL (for remote servers) or a command (for local servers). Claude connects to it, discovers what tools are available, and adds them to its toolset for that conversation.

Local vs remote connectors

There are two types of MCP server:

TypeHow it runsBest for
Local (stdio) A process on your own machine, launched by Claude Desktop Accessing local files, running local code, home automation
Remote (SSE/HTTP) A server at a URL, accessed over HTTPS Shared data, always-on services, mobile use, Claude.ai web

Local connectors are powerful but require Claude Desktop (not Claude.ai), and they only work on the machine where they're running. Remote connectors work everywhere — Claude.ai web, Claude Desktop, the mobile app — because they're just a URL.

For most users starting out, remote connectors are easier. You paste a URL and you're done. No installs, no config files, no local process to manage.

How to add an MCP connector to Claude.ai

Claude.ai (the web interface) supports remote connectors directly. Here's how to add one:

1

Go to Settings → Integrations

In Claude.ai, click your profile in the bottom left, then Settings. Find the Integrations tab (sometimes listed as "MCP" or "Connectors" depending on when you're reading this).

2

Add a custom integration

Click Add custom integration (or similar). You'll be asked for a name and a URL. The URL is the MCP server's endpoint — it typically ends in /mcp or /sse.

3

Authenticate (if required)

Most production MCP servers require authentication. The server will redirect you through an OAuth flow — usually Google or GitHub sign-in — and then return you to Claude with the connector active. This ties the connector to your account so your data stays separate from other users'.

4

Verify it's working

Start a new Claude conversation and ask it what tools are available, or call a tool directly. For a record store connector you might type usage() to see your current usage stats, or ask Claude to store something.

Claude Desktop users: Remote connectors work the same way — Settings → Developer → Edit Config, then add a JSON entry with "type": "sse" and your server URL. Local connectors use "type": "stdio" and specify the command to run.

What happens when a connector is active

Once a connector is added, Claude discovers its tools at the start of each conversation. If you have a record store connector, Claude might have tools like add(), search(), context(). These show up in Claude's tool menu.

Claude doesn't call tools automatically unless you ask it to (or unless the tool is framed as something that should run on startup). The tools just become available — you can use them explicitly or describe what you want and let Claude decide which tool to call.

The key benefit over pasting data directly: the tools return only what you asked for. A search over 50,000 records returns the matching few, not all 50,000. That keeps your context window clean and your costs down.

What makes a good MCP connector?

Not all MCP servers are equal. Things to look for:

The easiest remote connector to try: Stash

Stash is a hosted record store built specifically for Claude. It gives Claude a personal database: add records, search them, store standing context. The whole design is token-light — every response is structured to give Claude exactly what it needs without padding.

It takes about 90 seconds to set up:

  1. Copy the connector URL: https://app.stashlite.com/mcp
  2. Add it to Claude as a custom integration (Settings → Integrations)
  3. Sign in with Google when prompted — your account provisions automatically
  4. Ask Claude: context() to see your standing context, or add("your first note") to store something

You don't need to create an account separately. The OAuth step IS the signup. A free account gets 10,000 records and 100 queries per month — enough to run any of the major use cases (Notion offload, mini-CRM, daily context) without spending anything.

Try Stash — free, no setup

Connector URL:

https://app.stashlite.com/mcp Add to Claude →

Common questions

Does this work with Claude.ai Pro, or only with the API?

Remote MCP connectors work with Claude.ai (the web app) on most plans. You don't need API access. The connector URL is all you need.

Can I use multiple connectors at once?

Yes. Claude supports multiple active connectors simultaneously. Each one adds its tools to the session. Just be aware that each active connector adds a small overhead to your context window for the tool descriptions.

Is my data private?

With a properly-built remote connector, yes. Your data is stored under your account and the connector returns only your records to Claude. Stash uses Google OAuth for this — your records are scoped to your Google account. See the security page for specifics.

What if I want to build my own?

Anthropic has published the MCP spec and several SDKs (Python, TypeScript). If you want a simple data store, though, using an existing hosted one is faster — you're usually better off storing your data somewhere purpose-built than building your own persistence layer.

How is this different from Claude Projects?

Claude Projects is a conversation-level memory managed by Anthropic's UI — you paste documents in and they're available in that Project's context. MCP connectors are lower level: they're external tools your Claude instance can call, returning only the data it needs in real time. Projects are better for fixed documents; MCP connectors are better for large, searchable, dynamically updated data.

Getting more out of MCP

Once you've added a connector and got comfortable calling tools explicitly, the next step is using custom instructions to tell Claude to load your context automatically at the start of every conversation. With Stash you'd add something like:

At the start of every conversation, call context() to load my standing context.
Use search() to look up anything you're not sure about from my records.

That one instruction turns Stash into a silent background tool — Claude loads your context without you asking, and reaches into your records when it needs them. It's the simplest version of a personalised AI setup that actually works.

Ready to add your first MCP connector?

Stash is free to start. No credit card. Sign in with Google when you add the connector — that's all the setup there is.

Add Stash to Claude →

Related: Best Claude MCP servers — the practical list · Claude Projects vs MCP — what's the difference? · Claude persistent memory via MCP · Offloading Notion to Stash · Start-my-day with context()