Best MCP Servers in 2026: Complete Directory by Category

By 2026, the Model Context Protocol has settled into the default way to connect AI coding assistants to external systems. The original 2024 spec from Anthropic has been adopted across Claude Code, Codex CLI, and most other major AI dev tools. The result: a Cambrian explosion of MCP servers covering everything from databases to browser automation.

This guide is the curated, category-by-category answer to the question "which MCP servers should I actually install in 2026?" Every server here has been security-scanned by SkillsLLM's daily Semgrep + dependency audit pipeline, and the picks are sorted by community signal (GitHub stars, votes on SkillsLLM, real-world adoption).

How to choose an MCP server

Before installing anything, decide what you actually need the AI to do:

• Read-only data access (querying a database, reading from an API) - lowest risk, highest leverage. Start here.
• Mutating operations (writing files, sending messages, executing shell commands) - higher leverage, but requires trust. Audit the server's source code or rely on SkillsLLM's security scan results.
• Long-running processes (background workers, scheduled jobs) - rare, usually overkill for an interactive coding session.

The 100-star quality filter on SkillsLLM keeps low-effort scaffolding repos out of search results. If a server has under 100 stars, it is either too new to evaluate or too obscure to maintain.

Databases

The most-installed category by far. Every developer has a database; every database benefits from being queryable through natural language.

• Postgres MCP server - direct SQL access, schema inspection, query plan analysis. The starting point for anyone with a Postgres-backed app.
• SQLite MCP server - lightweight, file-based, perfect for local prototyping and one-off data analysis.
• MongoDB MCP server - document queries, index analysis, aggregation pipelines.
• Redis MCP server - cache inspection, key/value debugging, pub/sub monitoring.
• MySQL MCP server - same pattern as Postgres for the MySQL ecosystem.

Browse all database servers: /?category=mcp-servers and search for "database" or "sql".

Developer tools

Servers that connect the AI to your existing dev workflow. These are where the real productivity gains land.

• GitHub MCP server - issues, PRs, reviews, Actions runs, repo metadata. Probably the second-most-installed server after Postgres.
• GitLab MCP server - same pattern for GitLab-hosted teams.
• Linear MCP server - issue tracking, project status, sprint planning - read and write.
• Jira MCP server - for the enterprise-Jira reality.
• Docker MCP server - container inspection, log streaming, compose file management.

The GitHub server alone often pays for the cost of setting up MCP — it eliminates most of the "open the GitHub UI" friction in agentic workflows.

Productivity

These are the servers that make the AI useful for non-coding work — meeting notes, calendar wrangling, document editing.

• Google Drive MCP server - read and search documents, sheets, slides.
• Notion MCP server - read and write pages, query databases, manage workspaces.
• Slack MCP server - read channels, send messages, search history.
• Calendar MCP servers (Google Calendar, Outlook) - read availability, create events.

A note on permissions: productivity servers typically require OAuth scopes that grant broad access. Treat them like you would any third-party Slack app — review what they request before authorizing.

Search and retrieval

Augment the AI's reasoning with up-to-date external information.

• Brave Search MCP server - web search without Google's bot detection.
• Tavily MCP server - search optimized for LLM consumption (returns summaries, not just URLs).
• Perplexity MCP server - lets the AI delegate harder research questions to a stronger search backend.
• Context7 MCP server - up-to-date library documentation. Solves the "Claude's training cutoff is 6 months stale" problem for popular packages. See our Context7 vs Cursor Rules comparison.

For most agentic coding work, Context7 is the highest-leverage install in this category — it cuts hallucinated API calls dramatically.

Browser automation

The 2026 dark horse. Headless browser servers turn AI agents into general-purpose web automation tools.

• Playwright MCP server - the de facto standard. Full browser control, screenshots, accessibility tree extraction.
• Puppeteer MCP server - similar but tighter Chrome integration.
• Browserbase MCP server - cloud-hosted browsers when you need to run agents headlessly without a local Chromium.

These servers unlock workflows like "scrape the docs, fill out the form, take a screenshot" that previously required custom code.

Filesystem and shell

Lower-level building blocks. Most AI coding assistants ship some version of these built-in, but specialized servers add features.

• Filesystem MCP server - enhanced file ops with glob patterns, search across large repos, batch operations.
• Shell/Terminal MCP server - execute commands and stream output back. Higher risk; install with care.
• Memory MCP server - persistent key-value memory across agent sessions. Useful for long-running tasks.

How many MCP servers should you actually install?

Three to seven. More than that and your AI assistant's available-tool list gets noisy, which degrades tool selection accuracy. Pick:

1. One database server (whichever DB you actually use).
2. The GitHub server.
3. One search/retrieval server (Context7 or Tavily).
4. One productivity server if you do non-coding work in the AI (Notion, Slack).
5. One browser automation server if your workflow involves the web.

Add more only when a specific task demands it.

Security: what to check before installing

Every MCP server runs in your local environment with the privileges you grant it. Before installing:

• Check the security scan on the SkillsLLM skill page (PASS / WARNING / FAIL).
• Review the requested permissions if it requires OAuth or API keys.
• Read the install instructions in the source repo — never blindly run git clone && configure.
• Pin to a known-good commit for production use, rather than tracking main.

Read more about how SkillsLLM scans skills in our FAQ or browse the glossary for definitions.

What's next

The MCP ecosystem in 2026 is past the early-adopter phase but still moving fast. Expect:

• Tighter integration with editor extensions - VS Code, JetBrains, and Cursor all ship MCP-aware extensions now.
• Standardized auth - OAuth flows for MCP servers are still inconsistent; expect convergence.
• Per-tool capability discovery - clients learning what tools are most useful for a given task and dynamically loading.

Browse the full MCP Servers category for the complete current list, sorted by stars or community votes. New servers are added daily by the GitHub scraper.

If you build an MCP server, submit it — every submission gets a security scan and a category assignment within 24 hours.