20 Best MCP Servers to Install in 2026

In one sentence: MCP directories list thousands of servers, but only a small handful are worth keeping enabled long-term — install too many and your Agent gets dumber; pick the wrong ones and you're back to copy-pasting context by hand. If you already write code in Cursor, Claude Code, or Codex but still wonder whether to wire up GitHub or bolt on ten more SaaS connectors, this article filters dev tools, data & ops, collaboration SaaS, and search & memory into the 20 MCP Servers most worth installing in 2026 — with selection principles, config patterns, and common pitfalls.

Series reading: Claude Code MCP setup guide, MCP least-privilege permissions, CodeGraph MCP in five minutes.

Why MCP became the default interface for AI coding in 2026

Model Context Protocol (MCP) was open-sourced by Anthropic in late 2024, exploded in 2025 with Claude Desktop, Cursor, Windsurf, and similar clients, and by 2026 is the de facto Agent plugin standard: OpenAI Codex, GitHub Copilot, and VS Code extensions all support the same Server ecosystem. Its value isn't "another chatbot" — it's turning repos, docs, databases, browsers, and ticket systems into structured tools Agents can call, cutting down on manual context hauling like "can you look up PR #123 for me."

Compared to early custom Function Calling, MCP's edge is install once, reuse across clients: the same GitHub MCP config can serve both Claude Code and Cursor. Community and official Server counts passed a thousand in 2026, but quality varies wildly — this article only covers 20 with official backing or broad community adoption, active maintenance, and clear install docs.

Four selection principles: don't build a tool junkyard

Before the list, four principles the community keeps validating in 2026:

• Pick by workflow, not by leaderboard. If you ship React + GitHub PRs daily, Context7 + GitHub + Playwright covers 80% of your week. Notion MCP only earns its context slot when your team actually runs on Notion as a knowledge base.
• Cap total active tools. Cursor has a ~40 tool ceiling; Claude Code is more forgiving, but too many tools makes Agents pick wrong ones or burn tens of thousands of tokens upfront. In practice, 3–7 Servers and ≤30 tools total is the sweet spot.
• Prefer Remote HTTP slim builds. GitHub, Supabase, Linear, Sentry, and others ship official OAuth remote Servers with smaller tool sets than local Docker full installs — better suited for Cursor.
• Least-privilege by default. Read-only PATs, read-only database connection strings, Filesystem scoped to project dirs — see the permissions deep dive.

One diagram: workflow → MCP stack

Selection logic compresses to one causal chain: define daily tasks, map Servers, then constrain count and permissions.

Three starter stacks (by role)

Dev tools (8 servers)

These Servers act directly on your codebase and engineering pipeline — the main battlefield for most readers.

1. GitHub MCP (official)

What it does: Read repos and files, search code, manage Issues / PRs, check CI status, partial security scanning. Maintainer: GitHub official (github/github-mcp-server). Transport: Remote HTTP (OAuth) recommended; Cursor users should avoid the local full Docker build (a single Server can exceed 40 tools).

Best for: Any team hosting code on GitHub. Aligns with GitHub config in the Claude Code triple-connect — fine-grained read-only PAT or OAuth is enough.

2. Context7

What it does: Pull latest official docs and code examples by library name and version — dramatically cuts Agent API hallucination. Maintainer: Upstash community project; tops nearly every "best MCP" list in 2026. Transport: Remote HTTP or local stdio.

Best for: Anyone using third-party frameworks (Next.js, Tailwind, Prisma, SwiftUI…). Extremely high signal density — one Server that replaces ten.

3. Playwright MCP (Microsoft official)

What it does: Drive real Chromium/WebKit/Firefox, understand pages via the accessibility tree rather than screenshots, click, fill forms, assert. An order of magnitude more reliable than letting an Agent "guess if the UI looks right." Package: @playwright/mcp.

Best for: Frontend devs, E2E work, anyone who wants "change code, verify in browser immediately." Each run costs tokens and time — trigger on demand rather than leaving it always-on.

4. Fetch MCP (official reference)

What it does: Controlled HTTP GET, converts web pages to Markdown for the model. Package: @modelcontextprotocol/server-fetch. Good for docs, health checks, reading public API responses.

Note: Don't use it instead of Context7 for library docs; Fetch has limited support for dynamic SPAs — complex pages still need Playwright.

5. Filesystem MCP (official reference)

What it does: Read and write files within explicitly declared directories. Package: @modelcontextprotocol/server-filesystem.

Note: Cursor already has sandboxed file capabilities — often unnecessary. More valuable for Claude Code terminal workflows or cross-directory batch jobs. Always scope args root paths; never point at $HOME.

6. Git MCP (official reference)

What it does: Read git status, diff, log, branch info; some versions support controlled commits. Package: @modelcontextprotocol/server-git or mcp-server-git.

Best for: When you need the Agent to understand uncommitted changes and history without granting full shell access. Complements GitHub MCP: local repo vs. remote collaboration.

7. CodeGraph MCP

What it does: Local-index semantic code graph — symbol jump, dependency blast radius (codegraph_impact), cross-file relationships. ZavCloud blog benchmarks: on the same bug-fix task, shell invocations dropped from ~12 to ~5 after wiring it up.

Best for: Mid-to-large monorepos, fuzzy microservice boundaries, teams that need "change one place, know the impact." Deploy via the five-minute guide.

8. Serena

What it does: LSP-based semantic code search and edit, 20+ languages, understands symbols and references better than plain-text grep. Maintainer: Community (oraios/serena); rising reputation among heavy Claude Code users in 2026.

Best for: Multi-language monorepos needing precise rename / find references. Overlaps with CodeGraph — generally pick one or split by project.

Data & ops (4 servers)

9. Supabase MCP (official)

What it does: Query tables, run SQL, inspect schema, manage Edge Functions and branches (permission-dependent). Transport: Remote HTTP + OAuth.

Note: Production must use read-only roles or dev branches; writable mode mistakes are costly. Neon and other Postgres hosts offer similar MCPs — pick one for your stack.

10. DBHub / Bytebase (generic PostgreSQL)

What it does: One of Anthropic's recommended Postgres paths (original server-postgres is archived). Connect any PostgreSQL via DSN, run controlled queries. Install: npx -y @bytebase/dbhub --dsn "postgresql://...".

Best for: Self-hosted databases, non-Supabase stacks, teams wanting a unified SQL tool surface.

11. Sentry MCP (official)

What it does: Pull error events, stacks, releases, and issue status — lets Agents complete the first half of "error → locate → open PR" inside the editor. Transport: Remote HTTP.

Best for: Frontend and backend teams already on Sentry. Read-heavy; fixes still need GitHub + local debugging.

12. Docker MCP (community / Docker official preview)

What it does: List containers, read logs, controlled start/stop — helps Agents understand local or CI-simulated service topology.

Note: Large permission surface — dev machines only; production clusters need dedicated ops MCPs or read-only kubeconfig. Higher value if you containerize local dev entirely.

Collaboration SaaS (5 servers)

These Servers plug Agents into tools your team already pays for — only wire what you already use; don't switch ticket systems for MCP's sake.

13. Linear MCP (official)

What it does: Read/write issues, update status, search, align sprints. Remote OAuth, slim tool set.

Best for: Startups and product teams living in Linear. Paired with GitHub MCP for "ticket → branch → PR" end-to-end descriptions.

14. Notion MCP (official)

What it does: Search and read/write Notion pages and database views — PRDs, meeting notes, knowledge bases.

Note: Large workspaces can flood context; scope integration access to specific pages.

15. Slack MCP (official)

What it does: Read channel messages, post summaries, browse threads — sync Agent output into your collaboration flow.

Best for: Orgs where engineering communication runs through Slack. Keep write permissions off by default; human confirmation before posting.

16. Figma Dev Mode MCP (official)

What it does: Read structure, variables, and styles from selected Figma frames — supports design-to-code. Prerequisite: Desktop app open with a frame selected.

Best for: Design-system-driven frontend teams. Forms a triangle with Context7 (component docs) + Playwright (visual regression).

17. Stripe MCP (official)

What it does: Query customers, subscriptions, invoices, and test-mode data — assists billing feature dev and troubleshooting.

Note: Test keys or strict read-only only; payment compliance is high-stakes — never let Agents auto-execute production writes.

Search & memory (3 servers)

18. Exa MCP

What it does: Agent-optimized semantic web search — better than generic search for technical queries, papers, and docs. Requires API key.

Best for: Research tasks like "find the latest 2026 blog post / RFC / competitor implementation." Complements Fetch: Exa finds URLs, Fetch or Playwright reads deeply.

19. Brave Search MCP

What it does: Real-time web search via Brave Search API. Package: community @modelcontextprotocol/server-brave-search.

Best for: Users with existing Brave API quota wanting an Exa alternative. Pick one — don't duplicate tool slots.

20. Memory MCP (official reference)

What it does: Persist cross-session facts as a knowledge graph (project conventions, user preferences, architecture decisions). Package: @modelcontextprotocol/server-memory.

Best for: Solo devs or small teams maintaining one product long-term who want Agents to "remember" team norms. Unlike Sequential Thinking (honorable mention below): Memory stores facts; Sequential Thinking breaks down steps.

All 20 at a glance

Install & config examples

Config file locations differ by client, but JSON structure is nearly identical. Below are the two most common patterns in 2026.

Claude Code (~/.claude.json)

// ~/.claude.json → mcpServers
"context7": {
  "command": "npx",
  "args": ["-y", "@upstash/context7-mcp"]
},
"github": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-github"],
  "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxx" }
},
"playwright": {
  "command": "npx",
  "args": ["-y", "@playwright/mcp@latest"]
}

After saving, fully quit Claude Code, restart at the repo root, and run /mcp to verify. Full walkthrough: setup guide.

Cursor (~/.cursor/mcp.json or project .cursor/mcp.json)

// Same structure as Claude Code; for GitHub use Settings → MCP → add official Remote
// Avoid local github-mcp-server full install — 40+ tools can hit the ceiling
{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp"]
    },
    "playwright": {
      "command": "npx",
      "args": ["-y", "@playwright/mcp@latest"]
    }
  }
}

CodeGraph add-on (project-level)

# repo root
codegraph init -i
codegraph status
# add to ~/.claude.json:
# "codegraph": { "command": "codegraph", "args": ["mcp"] }

Common pitfalls

• "Tools don't show up after install" — JSON syntax error, client not fully restarted, or Claude Code not launched from repo root. Debug with /mcp.
• "GitHub MCP 401" — PAT missing repo scope, expired, or insufficient permissions; OAuth remote needs re-authorization.
• "Cursor tool limit exceeded" — Switch to GitHub Remote slim build, disable unused Servers, or turn off nonessential tools within a Server.
• "Agent still won't call MCP" — Prompt explicitly: "use GitHub MCP to read issue #N"; some tasks require allowing tool use. See smoke tests in the triple-connect overview.
• "Production data got modified" — Supabase / DBHub writable connection by mistake; Stripe live key misuse. Default to read-only + dev environments.

FAQ

Do MCP Servers overlap with Cursor's built-in capabilities? Partially: Cursor already has project indexing and terminal — Filesystem MCP is often skippable. But Context7, GitHub remote collaboration, and Playwright real-browser verification are still worth wiring externally.

Can Codex CLI use the same MCP stack? Yes. Codex and Claude Code both support the MCP protocol; config paths differ (~/.codex/config.toml or the client MCP panel), but Server entries are reusable.

Do I need all 20? No. The ⭐⭐⭐ trio fits almost everyone; pick the rest by stack, keeping 3–7 Servers active at once.

How do I know if MCP is worth it? Run the same task with and without MCP: count manual context copies and Agent detours before vs. one tool-call loop after. CodeGraph + GitHub triple-connect showed the clearest ROI in our benchmarks.

What should a security review cover? Each Server's OAuth scope / PAT permissions, Filesystem root paths, database DSN read/write roles. Checklist in MCP least-privilege permissions.

• Claude Code MCP setup guide
• MCP triple-connect overview
• MCP least-privilege permissions
• Deploy CodeGraph MCP on Cloud Mac in 5 minutes
• Claude Code vs Cursor (2026)