Fifteen MCP servers that earn a permanent slot in Claude Desktop, sorted by job: core tools, developer servers, work and docs connectors, and image and video generation. You get install steps for extensions, config files and remote connectors, plus the security rules that keep a growing setup safe.
Claude Desktop is a capable chat app on its own. Add the right MCP servers and it becomes the place where you read your files, review pull requests, fetch web pages and generate images without switching windows. The catch is that the ecosystem is now enormous, and installing twenty servers on day one is the fastest route to a slow, confusing and risky setup.
This list sticks to servers that earn a permanent slot. Every pick says what it does, where it fits and what to watch for. If you only have ten minutes, install the first two from the core section and add the rest when a real need shows up. Going into 2027, the winning habit is not collecting servers. It is choosing a few and trusting them.
💡 Quick answer: Start with Filesystem and Memory, add GitHub or Notion depending on your job, and put an image and video connector on top if you make visual content. Everything else is optional.
The Model Context Protocol (MCP) is an open standard that lets an AI app talk to outside tools through one consistent interface. Think of a universal port. Before MCP, every integration was custom wiring. Now a server written once can plug into any compatible client, and Claude Desktop is one of the most widely used of them.
An MCP server exposes up to three kinds of things:
When you ask "summarize the open pull requests on my repo," Claude picks the right tool, calls it, reads the result and answers in plain language. Claude Desktop asks for your permission before a tool runs, so you stay in control of each action. The practical result is a chat window that can act, not only talk.
Servers come in two flavors, and the difference shapes everything from install effort to risk.
| Type | Runs where | How you add it | Best for |
|---|---|---|---|
| Desktop extension (.mcpb) | Your computer | Double-click the bundle | Files, browser, local apps |
| Local server via config | Your computer | Edit a JSON file | Servers with no bundle yet |
| Remote connector | The vendor's cloud | Paste a URL and sign in | Hosted services like Notion |
Local servers can touch your disk and your apps. Remote connectors only touch what the vendor exposes behind a login. That one distinction drives most of the safety advice later in this article.
Claude Desktop currently offers two separate paths: local desktop extensions that run on your machine, and remote custom connectors added through the connectors menu. Pick the one that matches the server you want.
Open Settings, choose Developer and click Edit Config. That opens claude_desktop_config.json. On macOS it lives in ~/Library/Application Support/Claude/, and on Windows in %APPDATA%\Claude\. Here is a minimal entry for the Filesystem server:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/you/Documents/projects"
]
}
}
}
Quit Claude Desktop fully, reopen it, and the server's tools appear in the chat box menu. You need Node.js installed for any server launched with npx. Each extra folder or option goes in as another item in the args list.
💡 Debug tip: When a server refuses to start, the log files next to the config usually name the culprit. The usual suspects are a missing Node.js install, a mistyped path or a trailing comma in the JSON.
Open the connectors menu to browse the directory, an open catalog that mixes connectors verified by Anthropic with community entries. Desktop extensions install with a double-click on the bundle file. Older tutorials call these .dxt files, which is the legacy name for the same idea. Remote servers go in as custom connectors: paste the HTTPS address, then sign in on the vendor's own login page. No Node.js, no JSON.
Rule of thumb: use an extension or a connector when one exists, and edit JSON only when nothing else does. A bundle updates cleanly and a hand-edited entry is easy to break.
The reference servers live in the protocol's own repository: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. The maintainers describe them as reference implementations, so treat them as solid starting points rather than hardened enterprise software.
Filesystem reads, writes, searches and moves files, but only inside the folders you list. That boundary is the whole point. Use it to draft from a folder of notes, rename a batch of exports or let Claude refactor a small project without copy and paste. A good first request is "list every file in this folder changed this week and summarize what each one does."
Memory stores a small knowledge graph of people, projects and preferences that survives between chats. Tell Claude once that your reports use metric units and a formal tone, and it can recall that next week. Keep what you store boring and non-sensitive, because anything saved there is plain data on your machine.
Fetch downloads a web page and converts it to clean markdown so Claude can read documentation, release notes or an article you paste a link to. It is not a search engine. If you want search, add a dedicated search server such as the one Brave maintains for its search API.
Sequential Thinking gives Claude a structure for working through a problem in numbered steps, with room to revise earlier ones. It shines on planning questions, migrations and trade-off decisions. It also spends more tokens, so call it deliberately instead of leaving it on for every chat.
The Git and Time servers deserve a mention too. Git lets Claude read history and diffs in a local repository, and Time fixes the classic mistake of Claude guessing what day it is.
Code is where MCP pays back fastest. The picks below turn Claude from a code-writing chat into something that reads your repositories, tests your pages and checks live errors.
The official GitHub server gives Claude access to repositories, issues, pull requests and code search. Ask it to list stale branches, draft a changelog from merged pull requests, or review a diff against the issue it claims to fix. Create the access token with the narrowest scope that works, and begin with read-only repositories before you let Claude open or comment on anything.
Microsoft's Playwright server drives a real browser through structured page snapshots instead of screenshots, which is faster and lighter on tokens. Claude can open a page, click through a flow, fill a form and report what broke. It is the best option here for QA and smoke tests. It is slower than a direct API call, so reach for it when a page has no API.
Context7, from Upstash, pulls current, version-specific documentation for libraries into the conversation, which ends the habit of Claude confidently using an API that changed two releases ago. Sentry runs a hosted server, so Claude can read error groups and stack traces, then suggest a fix in the same chat. For databases, pick a maintained Postgres server and connect it with a read-only role.
Here is how the four fit together on a normal bug report. Sentry shows the failing line. GitHub shows the commit that touched it. Context7 confirms how the library is supposed to behave in your installed version. Playwright then reproduces the fix in a browser. That is four short tool calls in one conversation, with you approving each step.
| Server | Maintained by | Best for |
|---|---|---|
| GitHub | GitHub | Pull requests, issues, code search |
| Playwright | Microsoft | Browser testing and form flows |
| Context7 | Upstash | Fresh library documentation |
| Sentry | Sentry | Error triage and stack traces |
Not everyone writes code. These servers bring the tools of a normal workday into the chat, so the summary, the ticket and the draft all happen in one place.
The official Notion server lets Claude search pages, read databases and create new docs, so meeting notes can become a project brief without copying text around. Linear offers a hosted server for creating issues, updating their status and summarizing a sprint. Both work as remote connectors, so you sign in once and skip local installs.
Try a request like "read the planning page in Notion, then open a Linear issue for each unowned task." Claude reads, proposes the list and waits for your approval before it creates anything.
Figma's Dev Mode server hands Claude layout and style context from a selected frame, which makes generated front-end code far closer to the design. Zapier MCP exposes thousands of apps through a single connector. It is handy for the odd tool nobody has a dedicated server for, but it widens what Claude can touch, so restrict the actions you enable and review them every few weeks.
Claude writes excellent prompts, but it cannot draw. A generation connector closes that gap. The PicassoIA connector exposes four models to Claude: an image model, an image editor, a video model and a video model with sound. Jobs run asynchronously. Claude submits the request, checks until it finishes and hands you the result link. Up to five predictions run at once per account, shared across your connections.
| Model | Output | Use it for |
|---|---|---|
| Picasso IA Image | Image | Blog headers, product mockups, ad visuals |
| Image Editor Pro | Edited image | Retouching and object replacement |
| Picasso IA Video | Video | Short clips from text or a still image |
| Seedance 2.5 Lite | Video with audio | Social clips with sound |
This pairs well with the work servers. Draft a post in Notion, ask Claude to write a visual prompt for the header, generate it through the connector and paste the link back into the page. One chat, no tab juggling.
Try one request: "Write a visual prompt for a ceramic mug on a wooden table in morning light, generate the image, then animate it into a five-second clip." Claude drafts the prompt, submits the image job, polls until the file is ready, then feeds that image into the video model. You get two links and can ask for changes in the same chat: "warmer light," "slower camera move," "wider framing."
💡 Prompt tip: Ask Claude to name the lens, the light direction and the surface texture. Prompts with those details produce far more photographic results than a pile of vague adjectives.
The Claude family behind Claude Desktop is also in the Picasso IA catalog. Here is a quick way to test a model before you wire anything up:
Every server you add is either code running with your permissions or a vendor with access to your data. Treat it that way.
Give every server the minimum. Folder access for one project, not your whole home directory. A GitHub token limited to the repositories you need. A database role that can only read. Approve tool calls one at a time at first, and switch to "always allow" only for tools you have watched behave. Read the permission prompt before you click Allow. It names the tool, and you can expand it to see what Claude plans to send, which is your last chance to catch a wrong folder or an odd request.
Never paste a token into the chat. Put it in the env block of the config, or use a connector that signs in through OAuth. Keep the config file out of any repository you push.
Watch for prompt injection too: a web page or document Claude reads can contain text that tries to give it orders. A Fetch or Playwright session on an untrusted page, combined with write access to your files, is a risky mix. When you stop using a server, delete it. Fewer servers means fewer things that can go wrong, and fewer tool descriptions eating into the model's context.
💡 Monthly audit: Open the config and the connectors list once a month and remove anything you did not use. It takes two minutes.
You do not need every server in this article. You need the three or four that match your day. Use this table as a starting point:
| Role | Install first | Add later |
|---|---|---|
| Writer | Filesystem, Memory, Fetch | Notion, an image connector |
| Developer | Filesystem, GitHub, Context7 | Playwright, Sentry |
| Designer | Filesystem, Figma, an image connector | Zapier, Notion |
| Founder | Notion, Linear, Memory | Zapier, a video connector |
Then try the creative side. Open Picasso IA, pick Picasso IA Image for a header or product shot, and turn the result into a short clip with Seedance 2.5 Lite or Picasso IA Video. Generate one image and one clip, then compare them with what you would pull from a stock site. Seeing a real result from your own prompt is the fastest way to decide which servers deserve a permanent place in your setup. Revisit the table after a month of real use: the servers you reach for without thinking are the ones to keep.
