Best MCP Servers Today: Official, Hosted, and Community

The best MCP servers are the ones that fit your task, expose only the access it needs, and come from a maintainer who's still around. Start with the official reference servers or the vendor's own server, then decide whether to run it yourself or use a hosted one. This guide covers what's worth using in September 2026.

How to judge a "best" MCP server

Star counts and directory rankings tell you what's popular. They don't tell you what's good. Ask four questions instead:

• Is it maintained? Look at the last commit and the open issues. A server can be popular and abandoned at the same time.
• Can you turn off the dangerous tools? A server that mixes read-only lookups with delete and execute tools in one list is riskier to hand to an agent. You want writes to be a separate set you can disable.
• What does it need to log in? Some servers need nothing. Some need a browser sign-in. Database servers need a real password that has to live somewhere.
• Who runs the process? A local stdio server runs on your machine as a child of your AI app. A remote one runs at a URL, on someone else's infrastructure or your own.

If the terms are new, the MCP server glossary entry and What is MCP? cover the basics.

The official reference servers

The modelcontextprotocol/servers repository on GitHub is the most-cited "official" source. As of September 2026 it holds just seven servers, and they're reference and demo quality, not production tools:

• Everything: a test server that exercises every MCP feature.
• Fetch: fetches web content.
• Filesystem: reads and writes local files.
• Git: works with local repositories.
• Memory: a persistent, knowledge-graph memory store.
• Sequential Thinking: structured, step-by-step reasoning.
• Time: timezone conversion.

The TypeScript ones install with npx -y @modelcontextprotocol/server-<name>. The Python ones use uvx mcp-server-<name>.

That's shorter than most older roundups suggest. Thirteen servers, including GitHub, PostgreSQL, SQLite and Slack, were moved to a separate servers-archived repository. Archived doesn't mean maintained. If a list still links the old in-repo GitHub or PostgreSQL server, it's out of date.

Databases

For PostgreSQL, MCPifex hosts mcp-postgres-server. You enter your host, port, database and credentials once in the portal. query (read-only SELECT), list_schemas, list_tables and describe_table are on by default. The write tool execute is off until you switch it on.

The package also has a connect_db tool that would let a client point the connection at a different database. It isn't in the MCPifex catalog, so it's always refused. See the PostgreSQL server page for the full tool list, or how to query Postgres from Claude for a worked example.

If you self-host instead, pick a maintained implementation and use a database role with only the permissions you need.

Search and trend data

Two hosted servers cover search visibility and keyword research:

• Google Search Console has 19 tools for search analytics, URL inspection and sitemaps. You connect a Google account that owns the property or has full access. It can be a different account from your portal login.
• Google Trends has five read-only tools and needs no key: interest over time, related queries, trending now, regional interest and term comparison.

If you just need the raw text of a web page, the reference Fetch server does that.

Dev tools and version control

The reference Git and Filesystem servers work on local repositories and files. Restrict the directory they can reach, and remember that Git tools can change things, so don't treat the server as read-only.

For GitHub issues and pull requests, use the vendor's own github/github-mcp-server. It isn't the old archived reference server. It supports GitHub-hosted remote access and a local install.

Beyond the curated list: community registries

Lists like awesome-mcp-servers and directories like Glama and Smithery widen the search. Use them to find candidates, then read the actual repository. Check the release history, the docs for how it deploys, and what permissions it asks for. A big catalog doesn't tell you whether a particular server protects your data or has an owner who answers issues.

Who runs the process?

Most "best MCP servers" advice assumes you'll install everything locally with npx or uvx and keep credentials in your own environment. That's fine for a filesystem tool. It gets old fast for a server that holds a database password or an OAuth token, and every server you add multiplies the upkeep.

That's the gap we built MCPifex for. You create an instance of a server in the portal, save its credentials there, choose which tools are enabled (destructive ones start off), and get one API key for one endpoint, https://mcpifex.com/mcp. Test connection runs a real health-check call before you finish, and every call is logged per instance. The free plan includes 3 instances with no card.

Test any server before you rely on it

Whichever server you pick, run three quick checks:

1. Known answer. Ask for something you can verify, like the row count of a harmless table, and compare it with the source.
2. Out-of-scope request. Ask for something the server shouldn't do, and confirm it's refused.
3. Revoked access. Revoke the test credential and see what the client reports.

Write down the server version, client version, enabled tools and date. A working connection today says nothing about the next package upgrade, so rerun the checks after any change. For a team, name an owner for that list and give people a way to request a new tool instead of quietly adding one.

To connect, follow the guide for your client: Claude Code, Claude Desktop or ChatGPT.