Cursor's MCP support has matured enough that the question now is "which servers should I actually install?" - not "is this worth setting up?". This post answers that.
We've been running Cursor day-in, day-out across product engineering, infra work, and support - these are the servers that earn their place. Year-stamped because the MCP ecosystem moves fast and "best for 2026" is a moving target.
Why MCP in Cursor matters
Cursor without MCP servers is Cursor as the editor + Claude as the chat. Useful, but the chat lives in a window of context. The moment you ask "what's in this Postgres table?" or "comment on PR #123" the assistant has to guess.
With MCP servers, you give the AI a defined surface to do things. Read a file, query a database, comment on a PR, search the web, write to your filesystem. The chat becomes an agent that operates on your real systems.
The trade-off: every server you add is a small process the AI can call. You want servers that are useful for the work you actually do, not ten servers you'll check once and forget.
Below, the seven we install on every workstation, ranked roughly by how often they're called.
1. GitHub MCP
Why install it: Code review. Period. Read PRs, look up issue history, comment on diffs, scan a repo before context-switching to it. If your work touches GitHub, this is the first server to install.
Install (with 1Server):
Go to the marketplace, click Install on the GitHub server, paste your GitHub PAT once. The PAT goes into the encrypted vault, not into a config file.
Without 1Server you'd add this to your ~/.cursor/mcp.json:
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxx" }
}
}
}
Pro tip: Use a fine-grained PAT scoped to specific repos. The full repo scope is overkill for most work and a footgun if your AI ever decides to be helpful in unexpected ways.
2. Filesystem MCP
Why install it: Local file operations. Read, write, refactor, run tests. The filesystem server is what turns the AI from "smart commenter" into "agent that ships changes".
The official one ships with @modelcontextprotocol/server-filesystem and lets you scope access to specific directories - important for safety.
Install pattern:
{
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/code"]
}
}
Pro tip: Scope to your actual code root, not /. The first time the AI helpfully traverses your ~/Downloads you'll wish you had.
3. Postgres MCP
Why install it: Schema introspection + query exploration without leaving chat. "What's in the orders table" / "show me the last ten failed payments" / "what's the schema of analytics_events" - these are constant questions. Postgres MCP turns them into one-shot answers.
Auth: A read-only Postgres user with SELECT on the schemas you want. Don't give the AI write access to production. (Don't give the AI write access to staging either, in our experience.)
4. Web Search MCP (Brave or Perplexity)
Why install it: Up-to-date docs. The AI's training cutoff is months behind the npm release of every framework you use. A web search server fills that gap - "what's the new Tanstack Query pagination API" / "did this React RFC ship" / "what changed in Bun 1.2".
We use Brave Search MCP because the API is the cheapest of the searchable options and the results are fine for documentation queries. Perplexity MCP gives better synthesis if you can stomach their pricing.
5. Sequential Thinking MCP
Why install it: This one's an Anthropic-built reasoning aid that gives the AI a think tool. The AI uses it to break down complex problems before answering.
It sounds gimmicky and it isn't - for any non-trivial debugging or architecture question, the difference between Claude with Sequential Thinking and Claude without is real. You can see it; the answers get sharper.
It has zero side effects (no env vars, no auth, no external calls), so the install cost is zero.
6. Slack MCP
Why install it (optional): If your team lives in Slack, having the AI read recent messages from a specific channel ("what did infra say about the deploy yesterday?") is a force multiplier for context. If your team doesn't, skip this - it's pure overhead.
Auth: A Slack bot token scoped to specific channels.
7. Memory / Knowledge Graph MCP
Why install it: Persistent context across sessions. The Memory MCP server gives the AI a tiny knowledge graph it can write to and recall later. Useful for "remember that this project uses Tailwind v4 not v3" type sticky notes that the AI can self-update.
It's not magic - you have to actively use it for it to matter. But if you find yourself re-explaining the same project context every session, it's worth ten minutes of setup.
How to install all of these in 60 seconds
Manual setup with seven servers means seven mcpServers entries, seven secrets to paste into JSON, and a Cursor restart per change. With 1Server you do it once:
- Sign up at 1server.ai (free).
- Install the seven servers above from the marketplace, one click each.
- Create a 1Server API key.
- Add one entry to your
~/.cursor/mcp.json:
{
"mcpServers": {
"1server": {
"command": "npx",
"args": ["-y", "1server-mcp-engine"],
"env": { "ONESERVER_API_KEY": "your-api-key" }
}
}
}
- Restart Cursor once.
After that, adding or removing servers happens in the 1Server web UI; the Cursor config doesn't change.
What we left out
A few servers that come up in "best of" lists but didn't make our cut:
Browser automation MCPs (Playwright, Puppeteer, Chrome DevTools). Powerful, but most developers don't need automated browsing day-to-day. Install if you do QA or web scraping; skip otherwise.
Database migration MCPs. Most of them give the AI the ability to run schema changes, which is the kind of capability you want to grant carefully. We prefer reading-only Postgres + a separate, deliberate human-in-the-loop step for migrations.
Email and calendar MCPs. Cool in theory, surprisingly fiddly in practice - the auth surface is annoying (OAuth flows for Gmail / Google Calendar) and the AI's instinct to "just write the email" is something you have to manage.
These aren't bad servers, just not first-install material.
Further reading
- Cursor MCP setup guide - every detail of getting MCP working in Cursor specifically
- What is the Model Context Protocol? - primer if you're new to MCP
- 1Server vs Smithery - pick your tooling
If you're set up with three or four of these, you're already in a different productivity tier. The compounding curve on MCP is steep - every new server makes the next ones more valuable, because the AI can chain them. Install the GitHub one tonight; install Postgres tomorrow.