The 10 MCP Servers I Actually Use Every Day

I’ve been using Claude Desktop with MCP (Model Context Protocol) servers for almost six months now. I went through a phase of adding every server I could find to my config, then cut almost all of them out because they just added bloat, slowed down my AI, and rarely saved me any time. These 10 are the ones that stuck around— I open them every single work day, they solve real problems for me, and the tradeoffs are worth putting up with their flaws.

---

1. Official Filesystem MCP

What it’s best at: This is my most-used MCP by far. It lets my local AI read, edit, and create local files directly, without me copying and pasting hundreds of lines of code into the chat window every time I need a refactor. I use it for everything from editing markdown notes to refactoring React components in my side projects. It only shares the files you explicitly allow, so I don’t have to worry about accidentally uploading sensitive files to a third-party cloud, which is a huge win over AI tools that require you to upload everything.

One thing it’s bad at: Permissions are a constant headache on macOS, it can’t access any protected folders (like Documents or Downloads) without you manually granting full disk access in system settings, and it chokes on files larger than ~100MB. Worse, it follows symlinks by default, which is how I got into trouble. When I first set this up, I was lazy and added my entire home directory as an allowed path, instead of just specific project folders. I had a symlink from my projects folder to my `~/.ssh` directory, so I didn’t have to copy my keys over, and when I asked Claude to clean up unused files in projects, it followed the symlink, decided my 4-year-old RSA key was unused, and deleted it. I also lost a two-month-old todo.txt that was sitting in a folder I never meant to share. Now I only allow specific folders, and I check for symlinks every time I add a new path.

When would I replace it: If I worked primarily on remote servers instead of local projects, I’d swap this for the SSH MCP server that lets AI interact with files on the remote machine directly. If I needed stricter change tracking for every edit AI makes, I’d swap it for a git-integrated MCP that automatically commits every change, so I can roll back easier than I can with this.

My working config snippet for Claude Desktop:

```json

{

"mcpServers": {

"filesystem": {

"command": "npx",

"args": [

"-y",

"@modelcontextprotocol/server-filesystem",

"/Users/me/projects",

"/Users/me/notes/personal",

"/Users/me/notes/work"

]

}

}

}

```

---

2. Brave Search MCP

What it’s best at: Getting up-to-date public information without paying for a Google API key. I use this every day to check the latest version of an npm package, find release notes for a new tool I’m trying, debug error messages that are too new for my model’s training cut, and even check movie times when I’m going out. It returns full text snippets from results, so the AI can synthesize answers without having to click through every link, which is faster than the built-in browsing tool Claude has.

One thing it’s bad at: It can’t get past paywalls, obviously, and it’s worse than Google at niche local search (like finding a specific bakery that opened last month in a small town I’m visiting). It also pulls way too many irrelevant results if your query is vague, which can eat up a lot of context window for no reason.

When would I replace it: If I didn’t mind paying a few dollars a month for better results, I’d swap it for the Google Custom Search MCP, which has more relevant results for most queries. If I needed to search internal company docs behind SSO, I’d swap it for my organization’s internal search MCP.

---

3. Official Git MCP

What it’s best at: Walking through commit history, reviewing diffs, writing commit messages, and creating new branches without me leaving the chat. I use it every day when I’m debugging to figure out why a change was made six months ago, or when I can’t remember the exact git rebase command I need for my current branch. It’s way faster than me digging through the git log on the CLI myself.

One thing it’s bad at: It can’t resolve merge conflicts reliably, and it bogs down completely on large repos with thousands of commits, pulling way too much context and slowing Claude to a crawl. I’ve also had it commit directly to main when I forgot to explicitly tell it to create a new branch, which is a messy mistake I’ve had to fix more than once.

When would I replace it: If I needed to open PRs or add comments to GitHub/GitLab, I’d swap this for the official GitHub MCP, which handles that natively. If I worked with a large monorepo every day, I’d swap it for a lighter MCP that only interacts with the git API instead of pulling the whole local history.

---

4. SQLite MCP

What it’s best at: Querying and editing small local SQLite databases for personal projects. I run a personal finance tracker and a reading list that are both stored in SQLite, and I use this every day to get answers like “how much did I spend on coffee last month” or “what was the last fantasy book I read” without opening up a CLI or GUI client. It also helps me write new queries when I want to add a new report, fixing my syntax mistakes automatically.

You can spin it up locally with this simple command:

```bash

npx @modelcontextprotocol/server-sqlite /Users/me/personal-finance.db

```

One thing it’s bad at: It doesn’t handle write operations reliably. If the chat disconnects mid-write, it leaves an open connection that locks the database for hours until I manually kill the process. It also doesn’t support encrypted SQLite out of the box, which is a problem for me since I encrypt my finance db.

When would I replace it: If I needed encryption, I’d swap it for the SQLCipher-enabled fork of this MCP. If my db grew over 10GB, I’d swap it for the DuckDB MCP, which handles much larger datasets way better.

---

5. Postgres MCP

What it’s best at: Connecting to my local development Postgres databases when I’m building full stack apps. It automatically pulls table schemas, so I don’t have to keep switching to psql to check what column names I used, and it can write complex join queries for me when I need to pull data for debugging. I use it multiple times a week to answer quick questions like “how many users created an account in the last 7 days” without writing the query myself.

One thing it’s bad at: It doesn’t enforce read-only access by default, which is dangerous. I once connected it to my client’s staging database to pull some data, and I accidentally let Claude run an update query that messed up half the test users. I never connect it to anything that’s not local anymore. It’s also terrible at writing migration SQL, it forgets indexes and uses wrong data types more often than not.

When would I replace it: If I used Supabase for my projects, I’d swap it for the official Supabase MCP, which enforces read-only access by default and handles auth way better. If I used Prisma for my ORM, I’d swap it for the Prisma MCP that pulls schema from my local schema file instead of the database.

---

6. Puppeteer MCP

What it’s best at: Browsing dynamic websites for scraping and testing when there’s no public API available. I use it to pull menu data from local restaurant sites for a side project I’m building, and I use it to test my local dev server’s login flow to make sure it works after I make changes. It spins up a headless Chrome instance that Claude can control, so it can click buttons, fill out forms, and see content that loads with JavaScript.

My working config entry:

```json

"puppeteer": {

"command": "npx",

"args": ["-y", "puppeteer-mcp-server", "--headless=true"]

}

```

One thing it’s bad at: It eats RAM like crazy. I have 16GB of RAM on my laptop, and after a few hours of use, this server alone is using 1.5GB of memory. It also can’t get past most modern bot detection tools like Cloudflare Turnstile, so it fails immediately on a lot of popular sites.

When would I replace it: If I did a lot of scraping for work, I’d swap it for the Playwright MCP, which has better bot detection evasion and is more stable. If I only needed to fetch static content, I’d swap it for a simple HTTP fetch MCP that doesn’t spin up a whole browser and waste memory.

---

7. Official Jira MCP

What it’s best at: Pulling ticket details, updating status, adding work logs, and summarizing sprints without me switching over to the Jira tab in my browser. I do freelance work for a few clients that insist on Jira, so I use this every day to check what my next assigned ticket is, add my work log for the day, and write ticket descriptions when I need to open a new bug report. It saves me from having to navigate Jira’s clunky UI multiple times a day.

One thing it’s bad at: It times out constantly if you ask for more than 10 tickets at a time, so summarizing a whole 50-ticket backlog is out of the question. It also doesn’t handle custom fields well, which almost every Jira team uses. I’ve had it miss critical deadline information multiple times because it couldn’t read the custom “due date” field my client uses.

When would I replace it: If I had a choice, I’d swap Jira entirely and use Linear, so I’d use the Linear MCP which is way faster and more reliable. If my team used Asana or Trello, I’d swap for their respective MCP servers, all of which are lighter than this one.

---

8. Community Google Drive MCP

What it’s best at: Searching and reading my old Google Docs and Sheets without me downloading every file manually. I store all my client meeting notes and contracts in Drive, so I use this every day to pull the notes from our last meeting so I can remember what we agreed on, without opening Drive and searching through hundreds of files myself.

One thing it’s bad at: The setup is a nightmare. You have to create a Google Cloud project, set up OAuth consent, download credentials, and walk through a dozen steps just to get it working. It also randomly expires the auth token every couple of months, so I have to re-authenticate it all over again. It also can’t find files on shared drives half the time, even if they’re shared with me directly.

When would I replace it: If I used OneDrive for all my storage, I’d swap for the OneDrive MCP which has a way easier auth flow for personal accounts. If I stored all my notes in Notion, I’d swap for the Notion MCP which is way more reliable for searching notes.

---

9. Community Timezone MCP

What it’s best at: Converting time zones between me and my clients that are spread across 4 different time zones. I use this every single day to check what 2pm my time is in their time, and to make sure I’m not scheduling a meeting outside their business hours. It’s tiny, it uses almost no resources, and it does one simple thing really well.

One thing it’s bad at: It gets daylight saving time wrong for some smaller countries that change their DST dates often. I once scheduled a meeting with a client in Brazil an hour off because it had the wrong DST transition date, which was pretty embarrassing. It also doesn’t pull public holiday information, so I have to check that separately if I’m scheduling a meeting a month out.

When would I replace it: If I scheduled a lot of meetings and needed to check holidays and availability, I’d swap this for the Google Calendar MCP which includes all that functionality built in. If I worked with teams all over the world and needed more accurate DST data, I’d also swap it for the calendar MCP, which has more up-to-date data.

---

10. Wolfram Alpha MCP

What it’s best at: Accurate calculations, stats, and scientific data that general purpose AIs get wrong all the time. I use this every week for side project analytics, to calculate confidence intervals for A/B test results, convert weird units, and even help my niece with her college calculus homework. It’s way more accurate than Claude or GPT at raw math and factual scientific data.

One thing it’s bad at: The free tier only gives you 200 requests a month, which runs out really fast if you use it daily. You have to pay for a premium plan to get more requests, which is an extra expense I don’t love. It also just gives you the answer, it doesn’t explain the steps it took to get there, so it’s not great for learning if you need to see the work.

When would I replace it: If I didn’t want to pay for Wolfram, I’d swap it for the OpenAI Calculator MCP, which is free and works for simple calculations, even if it’s not as accurate for complex stuff. If I needed step-by-step explanations for math problems, I’d swap it for a community MCP that combines Wolfram with natural language explanation of the steps.

---

Actionable Next Steps

If you’re new to MCP or looking to clean up your current setup, here’s what I’d do next:

1. Open your MCP config right now and check your Filesystem MCP allowed paths. Remove any broad paths like your home folder, only add specific project and note folders, and check for symlinks pointing to sensitive directories like `.ssh` or your password manager folder. That one small step will save you from the same mistake I made deleting my private keys.
2. Don’t add all 10 of these at once. Every MCP server adds extra context your AI has to process, which slows down responses and uses up more of your context window. Start with just two: Filesystem for local work and Brave Search for up-to-date info, then add one new server a week that solves a specific pain point you have.
3. If you work with any non-local databases (even staging), create a read-only user for your database and only connect MCP with that user. Never use an admin or write-enabled user for any MCP connected to something that’s not your local dev machine.
4. Try one new MCP from this list this week that matches a problem you deal with every day. If you work with international clients, add the Timezone MCP. If you use local SQLite for side projects, give the SQLite MCP a spin.