You've seen "MCP" mentioned in Claude settings, on developer forums, and in AI tooling discussions. What actually is it?
Short version: MCP lets Claude reach out to external tools and services during a conversation. Instead of Claude only knowing what's in the chat window, it can query a database, read a file, call an API — and use those results to answer better.
Without MCP, Claude is isolated. Every conversation starts blank. Claude knows its training data, and it knows what you type into the chat. That's it.
That's fine for writing assistance and general reasoning. It breaks down when you need Claude to work with your specific data: your clients, your notes, your database, your documents.
The old solution was copy-pasting: you'd copy relevant context into the chat. This is slow, hits the context window limit fast, and means Claude only knows what you remembered to paste.
MCP is a better solution: a standard protocol that lets Claude call external services mid-conversation, get structured responses back, and use that data in its answers.
An MCP server is a small service that exposes "tools" — callable endpoints with defined inputs and outputs. Claude can see what tools are available and decide when to call them.
When Claude needs information from an MCP server, it makes a tool call. The server returns a structured result. Claude incorporates it into its response.
From your side, it looks seamless:
You: What do I know about the Acme project?
Claude: [internally: calls search("Acme", collection="projects") on your Stash MCP server]
Claude: Based on your records, Acme Corp is in the retail sector. The current project
is a Q3 strategy review with a July deadline. Last decision: Option B on pricing.
Would you like me to draft the exec summary?
Claude looked up your data. You didn't have to paste anything.
Run on your machine. Claude Desktop connects to them via a config file. Good for file system access, local databases, developer tools. Requires some setup and doesn't work on mobile or other devices.
Run on a server. You connect by pasting a URL into Claude settings. Works from any device, any Claude session. No local setup needed.
For most non-developers, hosted (remote) MCP servers are the practical choice. You get the capabilities without managing a local service.
The ecosystem is growing fast. Common categories:
| Category | What it enables | Example |
|---|---|---|
| Memory / notes | Store and retrieve your records | Stash — structured notes and lists |
| File access | Read and write local files | Filesystem MCP |
| Databases | Query databases of any kind | database connectors |
| Web tools | Search, fetch pages, scrape | Brave Search MCP |
| Dev tools | Run code, manage repos | GitHub MCP |
| Communication | Send emails, check calendar | Gmail MCP, Calendar MCP |
In Claude Desktop:
In Claude.ai (web, remote MCP):
For hosted MCP servers like Stash, you just paste a URL. No config file editing needed.
If you're new to MCP, start with a memory/notes server. It has an immediate, obvious payoff: you stop re-briefing Claude every session.
Stash is a hosted MCP that stores structured records. Add it to Claude in 30 seconds:
context() to confirm the connection worksThe free tier includes 2,500 records and 50 queries per month. Enough to get a feel for what structured MCP memory changes about your Claude workflow.