MCP & connectors2 min read
MCP Servers to Start With: What to Try and How to Choose
Official and reference MCP servers to connect AI with your files, GitHub, Notion, Stripe or Sentry, and how to tell whether a server can be trusted.
With MCP you can connect an AI assistant to your tools, but there are thousands of servers and not all are equally trustworthy. This is a selection to get started, separating reference servers (for learning) from each company’s official ones (for real work). To install them, follow how to install an MCP server.
Reference servers: for learning
The MCP project maintains a small set of servers that show how the protocol works. They’re ideal for experimenting locally:
| Server | What it does |
|---|---|
| Filesystem | Read and write files in the folders you allow |
| Fetch | Download web pages and convert them into a format models read well |
| Git | Read, search and manipulate git repositories |
| Memory | Persistent memory as a knowledge graph |
| Time | Current time and time-zone conversion |
| Sequential Thinking | Step-by-step problem solving with a sequence of thoughts |
| Everything | A demo with every protocol feature |
The repository warns they are reference implementations, not production-ready products. Use them to learn, in test folders.
Official servers: for real work
When you connect a real service, use the server published by the company itself. Some of the most useful:
| Service | For | Type |
|---|---|---|
| GitHub | Repositories, issues and pull requests | Remote, with a personal token |
| Notion | Search and edit pages and databases | Remote |
| Atlassian | Jira and Confluence | Remote |
| Stripe | Look up payments, customers and products | Remote |
| Sentry | Errors and performance of your apps | Remote, with OAuth |
Many apps also ship built-in connectors (Google Drive, Gmail, Slack, calendars…) that you turn on in settings without installing anything. If one exists, start there.
How to tell whether a server can be trusted
An MCP server can read your data and act on your behalf, so treat it like any program you install:
- Who publishes it? The service’s company or the official project beats an unknown author.
- Is it open source and maintained? Check when it was last updated and whether issues get answered.
- What permissions does it ask for? Only what its function justifies. A server to read your calendar doesn’t need to send emails.
- Remote or local? An official remote server avoids installing code on your computer; a local one gives you more control over the data.
- What external content does it read? If it reads websites, emails or third-party documents, it can receive prompt injections.
Good practice when connecting them
- Least-privilege tokens: read-only if you don’t need to write, and limited to the repositories or spaces needed.
- Few active connectors: each one adds tools the model must consider and possible risks.
- Review actions before approving them, especially those that write, delete or send.
- Disconnect what you don’t use.
If none does what you need, you can build your own MCP server in a short while.
Frequently asked questions
Where can I find more servers?
In the MCP Registry, the project's official directory, and in each app's connector directory (for example Claude's). Always prefer servers published by the service's own company.
Are the reference servers production-ready?
No. The repository itself warns they are reference implementations for learning and trying protocol features, and that everyone should assess their security before serious use.
