The Model Context Protocol (MCP) ecosystem has exploded in the past year. There are now hundreds of public MCP servers covering everything from databases to design tools to entire SaaS platforms — and most of them install with a single npx line.
This curated list highlights 20 open-source MCP servers that are genuinely worth setting up in 2026. They are grouped by use case so you can pick the ones that match your workflow, and each entry includes the install snippet you can drop into your claude_desktop_config.json immediately.
Developer essentials
1. Filesystem
The official filesystem server — read, write, and search files inside directories you whitelist.
{
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/code"]
}
}
Why use it: the single most-installed MCP server in existence. Gives any AI assistant the ability to read your codebase.
2. GitHub
The official GitHub server — search issues, create PRs, read repo contents, manage branches.
{
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
}
}
Why use it: AI-driven triage, PR reviews, release notes generation.
3. Git (local repo)
Works with any local Git repository — log, diff, blame, branch operations without needing GitHub.
Why use it: explore commit history, write commit messages, debug "who changed this and why" without leaving the chat.
4. Postgres
Read-only Postgres access — schema introspection plus arbitrary SELECT queries.
{
"postgres": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-postgres",
"postgres://readonly:pwd@localhost:5432/mydb"
]
}
}
Why use it: ad-hoc analytics, schema exploration, debugging a misbehaving query — all in natural language.
5. SQLite
Same idea as the Postgres server, but for SQLite files. Useful for local data exploration.
Why use it: every test fixture, every shipped app, every analytics dump that uses SQLite suddenly becomes queryable by Claude.
Productivity & SaaS
6. Slack
Search messages, post replies, summarize channel activity.
Why use it: "summarize what was discussed in #engineering today" — instant value.
7. Linear
Issue tracking workflow: search tickets, create them, transition status.
Why use it: turn a chat about a bug into a ticket in three exchanges.
8. Notion
Read and write Notion pages and databases.
Why use it: your wiki becomes Claude-queryable. Brilliant for onboarding docs and runbooks.
9. Google Drive
List, read, and search files across your Drive.
Why use it: "summarize that PRD I shared last week" works without copying anything into the chat.
10. Google Calendar
Check availability, create events, find meeting slots.
Why use it: scheduling assistance becomes a one-line chat.
Data & search
11. Brave Search
Web search via the Brave Search API — no Google scraping.
{
"brave-search": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-brave-search"],
"env": { "BRAVE_API_KEY": "..." }
}
}
Why use it: fresh-information retrieval without leaving the chat. Free tier is generous.
12. Fetch
Makes HTTP requests and returns parsed content (Markdown, JSON, plain text).
Why use it: "summarize this article" with a URL — the model uses fetch to pull the page itself.
13. Memory
A persistent key-value store the model can write to across sessions.
Why use it: Claude can remember facts about your projects, preferences, and ongoing tasks between conversations.
14. Time
Returns current time in any timezone. Tiny, but answers the surprisingly hard question "what time is it in Tokyo right now?"
Why use it: scheduling, timestamping notes, working across timezones.
Web automation
15. Puppeteer
Headless Chromium control — navigate, screenshot, scrape, click.
Why use it: "take a screenshot of our staging site" or "verify the login flow renders correctly" in one prompt.
16. Playwright
Alternative to Puppeteer with broader browser support (Chromium, Firefox, WebKit).
Why use it: if you already use Playwright in your test suite, this is the natural pick.
Cloud & infrastructure
17. AWS
Wraps the AWS SDK. List buckets, check EC2 instance states, query CloudWatch, manage Lambda functions.
Why use it: "which EC2 instances in us-east-1 are tagged staging?" returns the answer without leaving Claude.
18. Cloudflare
Manage zones, DNS records, Workers, KV stores, R2 buckets.
Why use it: "add a DNS record for staging.example.com pointing to this IP" becomes a chat command.
19. Docker
List containers, inspect logs, restart services on your local Docker daemon.
Why use it: debug a misbehaving local container without juggling terminals.
Niche but excellent
20. Figma
Read design files, extract variables and component metadata, inspect frames.
Why use it: generate component code from a Figma URL, audit design-token usage, document a design system — all from chat.
A pattern across all 20
Notice that every entry above installs the same way: one entry under mcpServers in your config, mostly via npx. That uniformity is the whole point of MCP. Adding any of these 20 servers is a 30-second config edit and a Claude Desktop restart.
How to choose which ones to install
A practical heuristic:
| If you regularly… | Install |
|---|---|
| Edit files locally | Filesystem, Git |
| Work on a team repo | GitHub |
| Query a database to debug | Postgres or SQLite |
| Live in Slack or Linear | Their respective servers |
| Take meeting notes | Notion, Google Drive |
| Schedule across timezones | Google Calendar, Time |
| Look things up online | Brave Search, Fetch |
| Maintain context across chats | Memory |
| Test UIs | Puppeteer or Playwright |
| Run cloud infrastructure | AWS, Cloudflare |
| Debug local services | Docker |
| Work in design systems | Figma |
Most developers find a sweet spot of 4–8 servers active at a time. More than that, and the LLM has to scan a long tool list before deciding what to call, which slows the first response slightly.
Where to find more
The official MCP server registry at modelcontextprotocol.io lists hundreds more, including:
- Spotify, YouTube, Apple Music — media control
- Stripe, Plaid — payments and banking data
- Sentry, Datadog — observability and incident triage
- Kubernetes, Terraform — infrastructure as code
- Apple Notes, Things, Todoist — personal task managers
If a SaaS exists, somebody has probably built an MCP wrapper for it. Search the registry before writing your own.
Conclusion
The quickest way to feel the power of MCP is to install three or four of these servers, restart Claude Desktop, and just start using your AI assistant for the kind of work you would normally split across five browser tabs.
My starter pack for any developer: Filesystem + GitHub + Postgres + Brave Search. Four servers, fifteen minutes of setup, and Claude Desktop transforms from a chatbot into something that can read your code, your repo, your database, and the open web simultaneously.
The rest of the list is just a function of what you do every day. Pick the ones that match your work and ignore the rest.
Try it yourself
Once the Filesystem + GitHub + Postgres trio is installed, here is the kind of multi-server conversation you can have:
orders table migration and tell me which fields it touched.search_issues, get_pull_request, then describe_tableThe relevant PR is #1842 “add orders table with payment fields”, merged 2026-04-22. It introduced these columns in orders:•
id (bigint, PK)•
customer_id (bigint, FK)•
status (text)•
total (numeric)•
currency (text)•
payment_provider (text)•
created_at (timestamptz)The table currently holds 14,302 rows. Want me to summarise how status values are distributed?
Three separate MCP servers cooperating in one answer. None of them know about each other — Claude composed the workflow on the fly because each server exposed its tools at startup.