The largest directory alone lists over 9,800 MCP servers for GitHub, Notion, databases, browsers, file systems, and pretty much anything that is a tool or a data source. If you’re wondering what these servers actually do and which ones you need: that’s what this article is about.
The protocol behind them, the open standard AI apps use to talk to external services, has its own article: what MCP is as a protocol. This one stays on the practical side: which servers exist and how to connect them.
What an MCP server does
An MCP server is a service that offers tools to an AI. Each tool is a named capability with a clear job: search my notes, read this file, create a GitHub issue. The client, meaning the AI app (Claude, ChatGPT, Mistral Vibe, Cursor), receives the list of tools when it connects, and decides on its own when to call which one.
That’s the whole trick. The server doesn’t need to understand AI; it just answers requests. The model doesn’t need to understand the service; it only sees described tools. Which is why the same server works with every client that speaks the protocol.
Which servers exist
The directories are full, but almost everything falls into four groups:
- Service connectors. GitHub, Notion, Slack, Google Drive, databases. The server translates between the AI and the service’s API. Useful when you already work there.
- System tools. File system, terminal, browser automation (Playwright, for instance). Mostly interesting for developers building AI agents or letting AI write code.
- Search and retrieval. Web search, documentation lookup, scraping. They give the model current information that wasn’t in its training data.
- Memory servers. They store what you tell your AI and make it findable in later sessions. The most interesting group if repeating your context forever annoys you, and the one with the biggest quality differences.
For everyday use you don’t need ten servers. Most people are fine with one or two: one for work (GitHub, Notion) and a memory.
Well-known MCP servers at a glance
The most searched-for servers, each with the route its vendor recommends (as of November 2026):
- GitHub: GitHub’s official server reads repositories and code, manages issues and pull requests, and automates workflows. GitHub hosts it at
https://api.githubcopilot.com/mcp/; locally it runs as a Docker image. - Notion: Notion runs a hosted server at
https://mcp.notion.com/mcp, with OAuth sign-in. It lets the AI read and edit pages and databases in your workspace. How that differs from the built-in AI is covered in Notion AI: what it does and what it costs. - Obsidian: There is no official server from Obsidian. Community servers run locally: some read the vault folder directly, others run as a plugin inside Obsidian, which then has to be open. The options are compared in Connecting Obsidian and Claude.
- Jira and Confluence: Atlassian offers the hosted Rovo MCP Server at
https://mcp.atlassian.com/v2/mcp, with OAuth 2.1 sign-in. It covers Jira, Confluence and other Atlassian products such as Bitbucket. - Figma: There are two official servers. The hosted one at
https://mcp.figma.com/mcpis available on all plans; the local one runs in the desktop app and needs a paid Dev or Full seat. - Memory servers: The MCP project’s reference server stores people, relations and observations as a small knowledge graph in a local file. Hosted memories like Memexa also take in documents and scans and give the citation. Memory MCP servers compared covers the options; how to connect Memexa is in Set up Memexa in Claude.
How to connect a server
Setup follows the same pattern in every client: enter the address, sign in once:
- Claude (claude.ai and Desktop): Claude calls MCP servers connectors. Settings → Connectors → Add custom connector, paste the server’s address. Sign-in opens in the browser.
- ChatGPT: What ChatGPT used to call connectors now lives under Plugins. At chatgpt.com/plugins → Add → Add custom MCP server, enter a name and the server URL, then sign in via OAuth. This works on the Plus plan and on the free plan; we’ve checked both ourselves. The full walkthrough is in Give ChatGPT and Claude the same memory.
- Claude Code: one terminal command, e.g.
claude mcp add --transport http memexa https://brain.memexa.ai/mcp. - Cursor: one line in the MCP config file.
After that, the server shows up in the client’s tool list and the AI uses it on its own whenever a question calls for it. What this looks like for Memexa is on the setup page, with copy-paste configuration for every AI app.
Local or hosted?
Many servers in the directories are small programs you start yourself: via npx in a terminal, configured in a JSON file. That’s fine for developers and the most common hurdle for everyone else: the server only runs while your machine runs, you handle updates yourself, and on your phone it doesn’t exist.
Hosted servers invert this. They run as a service on the web, you connect with a URL and a browser sign-in (OAuth), and the same state is reachable from every device and every client. In exchange, you hand your data to someone else, which brings us to the real question.
What to look for when choosing
The protocol is open; the quality behind it is not regulated. Three questions separate the solid servers from the risky ones:
- Where does the data live? With local servers, on your machine. With hosted ones: which country, under which law? A server that doesn’t answer this has already answered it.
- What is the server allowed to do? A tool that can delete files or send mail deserves more suspicion than one that only reads. Don’t connect more permissions than the task needs.
- Can you leave? If the service disappears or you want to switch: is there a complete export in an open format?
Memexa’s answers: EU hosting, only you, one-click Markdown export.
Frequently asked questions
Do I need to know how to code?
For hosted servers with OAuth sign-in: no. Entering an address and signing in is all. Local servers from the community directories usually assume a terminal doesn’t scare you.
How many servers can I connect at once?
Technically many; practically, restraint pays off. Every server makes the tool list longer, and the model walks it on every request; too many similar tools make its choices worse, not better.
What separates memory servers from each other?
Mainly three things: whether they run locally or hosted, whether they take only text snippets or whole documents too, and whether they tell you where an answer came from. Memory MCP servers compared puts Basic Memory, Mem0, Supermemory, Memexa and the reference server side by side.