
Most MCP server lists only cover developer tools. Here are the servers that actually stay installed, grouped by the job they do: files, code, media generation and memory.

The official Notion server. Read, search and update pages and databases from inside your AI. Best for turning a conversation into a document without copy-paste. Setup guide.
Summarize channels, draft replies and post updates from chat. The official workspace app needs admin scopes; the community server is simpler. Setup guide.
The official Linear server: create and triage issues from a conversation. If your team lives in Linear, this one removes the most context switching. Setup guide.
GitHub's official server: search code, read files, review PRs, manage issues. The default choice for anything repo-related. Setup guide.
Search and read Drive files so the AI answers from your actual documents instead of guesses. Setup guide.
Fixes the most common AI coding failure: made-up APIs. It injects real, version-specific library docs into the model's context. If you install one dev server, make it this one.
Microsoft's browser automation server. Lets an agent actually click, fill forms and verify a page works, not just describe how. The most-mentioned dev MCP for a reason.
Google's server for inspecting pages: performance traces, console errors, DOM state. Pairs with Playwright when you need diagnosis rather than just clicking.
Most MCP servers read or write text. Krater's server is the odd one out: it gives any connected assistant real generation tools. Images, video, music, voice, documents and research, exposed as about 20 tools, backed by 500+ models (Sora, Veo, Kling, Flux and the rest) so your client calls one server and the right model handles the job.
It is a remote server: one URL, OAuth sign-in, works with Claude, ChatGPT, Gemini, Perplexity and any MCP client. Included with Krater Ultra and Max plans, running on the same plan credits. Krater MCP setup or see the per-assistant pages: Claude, ChatGPT, Gemini, Perplexity.
Honest caveat: because Anthropic's public connector directory does not accept media-generation connectors, Claude users add Krater as a custom connector rather than finding it in the directory. It takes the same two minutes.
Read-only database servers let an assistant answer questions straight from your data. Great for internal reporting; always use a read-only role.
The reference fetch server turns any URL into clean text the model can read. Simple, but it makes "summarize this link" actually work everywhere.
A knowledge-graph server that gives an assistant persistent memory between sessions. Useful in long-running projects where context resets hurt.
Every pick above installs the same way: add the server's URL or command in your client's connector settings, approve access, and ask. Krater's MCP page shows the flow for Claude, ChatGPT, Gemini and Perplexity in one place, and our connect guides walk through the per-app details.
The servers themselves are usually free or open source; you pay for what they access. A GitHub server uses your GitHub account, and a generation server like Krater's runs on your plan.
An MCP server can do anything its permissions allow, so treat it like installing an app: prefer official servers, read the scopes you approve, and disconnect what you stop using.
Yes: ChatGPT connects to remote MCP servers through its apps/connectors settings, so the same server can work in ChatGPT and Claude alike.
The server is the package you connect; the tools are the individual actions it exposes. One server, many tools.