The 20 best MCP servers in 2026 (that actually work)

Model Context Protocol (MCP) servers turn your AI agent into a power tool. These 20 are the ones we trust in production — what each does, setup time in minutes, and the security caveats that actually matter.

In 2026, the MCP ecosystem is mature enough that the question is no longer "should I use MCP" but "which servers should I install first." This is the shortlist we recommend to teams setting up a coding agent stack from scratch.

For background on what MCP actually is, see our deep-dive on Model Context Protocol (MCP).

The 20 best MCP servers at a glance

The "install first" trio (every team needs these)

If you install only three MCP servers, install these.

1. Filesystem MCP

Maintained by: Anthropic (official) Setup time: 1 minute What it does: Lets your agent read, write, and search files in a directory you whitelist.

This is the default expectation now — any coding agent that can't touch your filesystem feels broken. The official server has sandboxing (you whitelist directories) and respects gitignore patterns.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/your/project"]
    }
  }
}

2. Git MCP

Maintained by: Anthropic (official) Setup time: 1 minute What it does: Git operations — branches, commits, diffs, log, status.

Without this, your agent has to shell out to git commands and parse text output. With it, structured Git operations become a first-class capability.

3. GitHub MCP

Maintained by: GitHub (official) Setup time: 5 minutes (needs a PAT) What it does: Repos, issues, PRs, comments, reviews, GitHub Actions.

The killer combo with Cursor or Claude Code: your agent reads the issue, writes the code, opens the PR — all without leaving the editor.

Code & development MCP servers

Linear MCP

Maintained by: Linear (official) Setup time: 3 minutes What it does: Read issues, create tickets, update status, add comments.

The combo of Linear MCP + GitHub MCP is the closest thing to "ticket-to-PR autonomy" we have in 2026 without paying for Devin.

Sentry MCP

Maintained by: Sentry (official) Setup time: 3 minutes What it does: Read errors, group reports, recent issues, performance traces.

Your agent can pull the production error stack trace, navigate to the source file, and propose a fix — all in one session.

Playwright MCP

Maintained by: Microsoft (official) Setup time: 10 minutes What it does: Full browser automation — clicks, fills, screenshots, JavaScript execution.

The replacement for fragile Puppeteer scripts. Your agent can navigate real web apps the way a human would. Pair with a vision-capable model for the best results. See our glossary on browser agents.

Data & database MCP servers

Postgres MCP

Maintained by: Community (multiple options) Setup time: 5 minutes What it does: Run SQL queries against a Postgres database.

Game-changer for anyone using AI for data analysis. Pick a read-only variant for analytics workloads; reserve write-capable variants for trusted contexts only.

Security note: Always use a read-only connection user. Never give an agent direct write access to production databases.

SQLite MCP

Maintained by: Anthropic (official) Setup time: 1 minute What it does: Query and update local SQLite files.

Underrated. Great for prototyping data pipelines, exploring downloaded datasets, or letting an agent maintain a small local knowledge base.

Knowledge & docs MCP servers

Notion MCP

Maintained by: Notion (official) Setup time: 5 minutes What it does: Search pages, read content, write new pages, query databases.

If your team's docs live in Notion, this server is non-negotiable. Your coding agent can pull architectural decisions from the wiki and apply them to code without you copy-pasting.

Slack MCP

Maintained by: Community (Slack-blessed) Setup time: 5 minutes What it does: Read channels, search messages, send messages, react to posts.

Two patterns we like: (1) agent reads design discussions before implementing; (2) agent posts daily PR summary to a team channel.

Security caveat: Scope the token to specific channels. A Slack token with workspace-wide read access is a prompt injection vector.

Web & search MCP servers

Brave Search MCP

Maintained by: Anthropic (official) Setup time: 2 minutes (API key) What it does: Web search optimized for LLM consumption.

The default web-search MCP. Brave's API is privacy-respecting and returns cleaner results than scraped alternatives.

Fetch MCP

Maintained by: Anthropic (official) Setup time: 1 minute What it does: Fetch a URL and return it as clean markdown.

Pair with Brave Search MCP for a complete "research and read" loop. Your agent searches, picks results, fetches, summarizes.

Cloud & infrastructure MCP servers

AWS MCP

Maintained by: Community Setup time: 15 minutes (IAM setup is the slow part) What it does: EC2, S3, Lambda, RDS, CloudWatch operations.

Powerful and dangerous. Use read-only IAM roles by default. Move to write-capable only for specific automation tasks.

Stripe MCP

Maintained by: Stripe (official) Setup time: 5 minutes What it does: Customer lookups, charges, refunds, subscription management.

For agents handling support or finance ops. Always test with a restricted-key Stripe environment before production.

Workflow & glue MCP servers

Zapier MCP

Maintained by: Zapier (official) Setup time: 3 minutes What it does: Trigger Zapier workflows, which then connect to 7,000+ apps.

The "I want my agent to do X but X has no MCP server" escape hatch. Slower than direct MCP but covers the long tail of integrations.

Sequential Thinking MCP

Maintained by: Anthropic (official) Setup time: 1 minute What it does: Structured reasoning tool that helps the agent plan complex tasks step-by-step.

Not a data source — a reasoning helper. Worth installing for complex multi-step workflows.

How to install MCP servers (the 5-minute primer)

Most MCP servers install the same way across coding agents:

1. Find the config file for your agent. Cursor: ~/.cursor/mcp.json. Claude Code: ~/.claude.json. Cline: settings UI.
2. Add the server entry with the right command and arguments.
3. Restart the agent or reload the MCP config.
4. Verify by asking the agent "what MCP servers do you have access to?"

For each server above, the exact config is in their respective README. Copy-paste, swap in your tokens, restart.

Security checklist for MCP servers

Before installing any MCP server, especially community-maintained ones:

1. Read the source. MCP servers are small (usually < 500 lines). You can audit them in 10 minutes.
2. Scope tokens narrowly. Read-only by default. Per-channel/per-repo/per-project tokens beat workspace-wide.
3. Watch for prompt injection surfaces. Any MCP server that reads untrusted content (web pages, emails, PRs from external contributors) is an attack vector.
4. Pin versions. npx -y @vendor/server@latest is convenient but can pull a compromised update silently. Pin to a specific version for production use.
5. Audit logs. If your agent runtime supports audit logging, enable it for all tool calls.

What we excluded from the top 20

A few MCP servers that get attention but didn't make this list:

• OpenAI MCP — exists but mostly redundant when your agent already uses an LLM API
• HuggingFace MCP — too narrow; useful for ML teams but not general
• Generic database MCPs — Postgres/SQLite cover most needs; specialized DBs (Mongo, Redis) only matter for specific stacks
• Custom internal MCPs — the highest-value MCP servers in most orgs are the ones you build for your own internal tools. Worth building once you've installed the externals.

Related reading

• Model Context Protocol (MCP) — the protocol explained
• Best coding agents in 2026 — which agents to install MCP on
• Claude Code vs Codex CLI — terminal agents that benefit most from MCP
• AI agent framework — the broader ecosystem