Best MCP Servers Worth Installing in 2026

MCP servers are the practical payoff of the Model Context Protocol. After installing 30+ in various setups, here are the ones that actually pay back the setup time. If you’re new to MCP, start with What is MCP for the explainer.

TL;DR — start with two

For most developers:

1. Filesystem — Claude/Cursor reads your real project files
2. GitHub — repo operations, PR/issue context, code search across your repos

Add these as you need them: Postgres, Brave Search, Slack, Stripe, Puppeteer.

The eight servers at a glance

#	Server	Maintainer	Best for	Risk profile
1	Filesystem	Anthropic (official)	Reading your real project files	Medium (scope to specific dirs)
2	GitHub	Anthropic (official)	Issues, PRs, repo search	Medium (scope PAT)
3	Postgres	Anthropic (official)	Read-only DB queries	Low (use read-only credentials)
4	Brave Search	Brave / Anthropic	Live web search	Low (read-only)
5	Slack	Anthropic (official)	Channel summaries, posts	Medium (scope bot token)
6	Puppeteer	Community	Browser automation	High (be careful)
7	Memory	Anthropic (official)	Persistent cross-session memory	Low
8	Stripe	Stripe (first-party)	Charges, customers, subscriptions	High (use restricted key)

The eight servers worth knowing

1. Filesystem (@modelcontextprotocol/server-filesystem)

Reference implementation from Anthropic. Reads, writes, lists files in a directory you specify.

What it gives you: Claude or Cursor can actually look at your files instead of guessing. The single biggest quality-of-life improvement when you start using MCP.

Config:

"filesystem": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Projects/my-project"]
}

Safety note: point it at specific project directories, not your whole home folder. You can configure multiple filesystem servers for different projects.

2. GitHub (@modelcontextprotocol/server-github)

Reads your repos, files, issues, PRs. Creates PRs, comments, labels.

What it gives you: “look at the open issues in my repo and suggest which to prioritize” actually works. “Create a PR with these changes” works.

Requires: GitHub personal access token. Scope it to specific repos when possible.

Config:

"github": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-github"],
  "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
}

3. Postgres (@modelcontextprotocol/server-postgres)

Read-only Postgres queries from Claude.

What it gives you: “what tables are in my staging database?” “show me users created in the last 7 days.” Schema-aware queries without you having to remember table names.

Config: connection string in the args. Use a read-only role for safety.

Variants: SQLite server, MongoDB server, BigQuery server — all exist as separate MCP servers.

4. Brave Search (@modelcontextprotocol/server-brave-search)

Web search via Brave’s API (privacy-focused, decent quality, free tier).

What it gives you: Claude can search the live web in conversation. “What’s the latest version of FastAPI?” gets a real answer.

Requires: Brave Search API key (free tier covers personal use).

Alternatives: Google Custom Search MCP, Tavily search MCP, DuckDuckGo MCP.

5. Slack (@modelcontextprotocol/server-slack)

Reads channels, posts messages, searches Slack history.

What it gives you: “summarize what happened in #engineering this week” — real summaries from real channel content.

Requires: Slack bot token with appropriate scopes. Don’t grant broader scopes than necessary.

6. Puppeteer (@modelcontextprotocol/server-puppeteer)

Browser automation. Claude can navigate to URLs, screenshot pages, click elements, fill forms.

What it gives you: “go to this site and screenshot the pricing page.” “fill out this contact form with these details and submit.” Real browser-driven workflows from the chat interface.

Use cases: scraping, monitoring, lightweight QA, exploring documentation sites.

7. Memory (@modelcontextprotocol/server-memory)

Persistent memory across sessions. Stores facts, preferences, project context as a knowledge graph.

What it gives you: Claude remembers what you told it last week. “I prefer Python over TypeScript for data work” persists across conversations.

Honest take: less essential than the others but powerful once you have it running. The persistent context across sessions changes how you use the model.

8. Stripe (Stripe-maintained)

First-party from Stripe. Reads charges, customers, subscriptions, invoices. Creates checkout sessions.

What it gives you: “show me last week’s failed charges and group them by error type.” Real business intelligence from the chat interface.

Requires: Stripe API key. Restricted key with read-only scope is safer than the full key.

Niche picks worth knowing

These solve specific problems well:

• Sentry MCP — query error events, see issue trends. Great for debugging conversations.
• Linear MCP — read/create issues, search tickets. If you use Linear, this changes the planning workflow.
• Notion MCP — search Notion workspace, read pages, create entries. Underrated.
• Read-the-Docs MCP — fetches arbitrary RtD-hosted documentation. Saves the “copy this doc into chat” pattern.
• Vercel MCP — read deployments, logs, project state. Used during incident response.
• Cloudflare MCP — DNS, Workers, Pages operations.
• AWS MCP servers (community + AWS-maintained variants) — Boto3-backed; useful for ops conversations.

The four-server starter stack

For most developers writing software professionally in 2026:

1. Filesystem (your active project)
2. GitHub (your repos)
3. Brave Search (live web context)
4. Memory (persistent context)

That’s roughly 10 minutes of setup and meaningfully changes how you work with AI tools. Adding more is incremental from there.

Setup time honesty

The MCP setup experience in early-2025 was rough — manual config-file editing, server processes to debug, transport types to understand. By mid-2026, Cursor and Claude Desktop have added install UIs, so the experience is closer to “browse and install” than “edit JSON by hand.”

If you’re starting fresh in 2026:

• Claude Desktop: Settings → Developer → MCP shows installed servers and config
• Cursor: Settings → MCP shows the same plus server health
• Continue.dev / Cline / Zed: similar UIs

Manual config editing is now mostly a fallback for advanced cases.

Building your own

If you have an internal API, database, or tool that you want Claude/Cursor to access, building your own MCP server is straightforward. The reference SDKs (@modelcontextprotocol/sdk for Node, mcp for Python) handle the protocol; you write the tool implementations.

Real example: a 50-line TypeScript MCP server that wraps your internal CRM’s API gives Claude direct read access to customer records. The integration is yours to maintain; the LLM-client integration is standardized.

Where to install vs not

• Personal machine, individual developer: install everything you’d use day-to-day
• Team-shared environment: stick to vetted servers, prefer hosted variants when available
• CI / production code: don’t run MCP servers in CI; that’s not what they’re for. MCP is interactive, human-in-the-loop tooling.

What we’d watch (briefly)

• MCP hosted/marketplace services — paid hosting for trusted MCP servers; growing as a category
• Authentication standards — OAuth flows for MCP servers needing cloud credentials are stabilizing
• Multi-modal MCP servers — image generation servers, voice servers, video processing — early but emerging

Verdict

For most readers: start with the filesystem + GitHub combo, add Brave Search if you research a lot, add Postgres or similar if your work involves databases. The whole stack is free to run locally. The compound benefit shows up in week 2-3 when the pattern of “Claude has real context” becomes default.

For the broader workflow, see Claude Code Workflow Guide. For how MCP-enabled tools compare against each other, see Claude Code vs Codex CLI vs Cursor Agent.