MCP and Connectors: Wiring Your Own Tools and Data to Claude
Out of the box, Claude only knows what it was trained on plus whatever you paste into the chat. Connectors and MCP are how you change that — how you let…
Claude guide · as of July 10, 2026 · 5 minutes read · details change — confirm current specs on claude.ai
Out of the box, Claude only knows what it was trained on plus whatever you paste into the chat. Connectors and MCP are how you change that — how you let Claude read your calendar, search your files, query a database, or take actions in another app. This guide covers what that actually enables, how much setup it really takes, and the safety questions worth asking before you plug anything in.
What MCP actually is
MCP stands for Model Context Protocol. Anthropic introduced it in November 2024 as an open standard — not a Claude-only feature — and other AI tools have since adopted it. The plain-English version: MCP is a common plug shape, a bit like USB-C for AI. Instead of every tool inventing its own way to talk to a model, MCP gives them one shared way to expose data and actions.
Connectors are the friendlier, packaged version of the same idea inside the Claude apps. A connector is usually an MCP server that someone has already built and hosted, so you sign in and grant access rather than running anything yourself.
What it lets you do
Once connected, Claude can pull in live context and, in some cases, act:
- Search and summarize files in a linked drive or document store.
- Read your calendar or tickets to answer "what's on my plate."
- Query a database or an internal tool in plain English.
- In Claude Code, wire up developer tools so Claude can run tasks against a real project.
The exact connectors available — and which plan and surface expose them — change often. Check the current connector directory in the Claude apps and the docs at anthropic.com rather than trusting a list from any guide.
Three ways to give Claude your data
| Approach | Setup effort | Best for | Watch out for |
|---|---|---|---|
| Directory connectors | Low — click, sign in | Popular apps: drives, calendars, dev tools | You grant real account access |
| Custom / self-hosted MCP server | High — config file, run a server | Your own database or niche/internal tools | You own the security and upkeep |
| Upload files or use Projects | Very low | One-off docs, static reference | A snapshot, not live data |
If you only need Claude to read a handful of documents, plain file upload or a Project is simpler and safer than a connector. Reach for MCP when you need live or interactive access.
The setup reality
There are two honest tiers of effort. Directory connectors are genuinely easy: pick one, sign in through the app's normal login (often OAuth), approve the scopes, done. Custom MCP servers are a developer task — you edit a config file, then either run a small server on your own machine or point Claude at a remote server URL. The Claude desktop app and Claude Code are the main places people wire up custom servers today. If a "5-minute setup" involves pasting a config block and a command line, that is the custom tier — budget more than five minutes.
What breaks
Connectors add power and new failure modes at the same time:
- Prompt injection. If Claude reads a document or web page that contains hidden instructions, those instructions can try to hijack what it does with your connected tools. This is the big one, and it is not fully solved.
- Over-broad permissions. Many connectors ask for more access than the task needs. A read task should not hold write-and-delete rights.
- Confidently wrong actions. Claude can be wrong. When a tool can send, pay, or delete, a wrong call has real consequences — keep those behind confirmation.
- Silent breakage. Third-party and self-hosted servers go stale, change APIs, or return junk. Claude may not flag it clearly.
- Context bloat. Pulling in huge results eats your context window and can bury the actual question.
Safety questions to ask first
Before you connect anything, walk this list:
- What exact permissions am I granting — read only, or write, send, delete?
- Who built and hosts this server: Anthropic, a known vendor, or a stranger's repo?
- Can I scope it to one folder or project instead of my whole account?
- Can I revoke access quickly, and do I know where?
- Is any of this data I would regret sending to a cloud model?
That last point matters: a local MCP server running on your machine still feeds its results to Claude in the cloud. Running the server locally does not make the request local, and turning off training does not either. If data cannot leave your device, MCP to Claude is not the answer — a local model is.
When this is the wrong tool
- For truly private data that cannot go to a cloud service — use a local model instead.
- For a one-time question about a few files — upload them or use a Project; skip the wiring.
- For unattended actions with real stakes (money, mass deletes, outbound messages) — a confidently-wrong model plus broad permissions is a bad trade. Keep a human in the loop.
- If you cannot answer the safety questions above — do not connect it yet.
Used deliberately, MCP turns Claude from a clever chat box into something that works with your actual stuff. The discipline is narrow scopes, trusted servers, and a human on the risky actions.
Related
- Integrations — the broader picture of connecting Claude to other apps.
- Claude Code — where developers wire up MCP servers most.
- Privacy — what leaves your device and what does not.
