Best MCP servers in 2026: a curated list with setup commands

The best MCP servers for most developers are GitHub, Playwright, Context7 and the error tracker or database you already use, such as Sentry or Supabase. This list groups useful servers by job, gives the exact command to add each one to Claude Code, and explains how to keep them from leaking data or flooding your context.

What are the best MCP servers in 2026?

The best MCP servers are the ones that connect your coding agent to the systems you touch every day: your code host, a browser, current library docs, your error tracker and your database. For most developers that means GitHub, Playwright, Context7, Sentry and Supabase (or your own database), plus one project tracker such as Linear.

If MCP is new to you, the short version: the Model Context Protocol is an open standard that lets AI applications call external tools. An MCP server exposes those tools, and a client such as Claude Code, Codex or Cursor connects to it. Our explainer on what MCP is covers the architecture.

Here is the shortlist this guide covers, grouped by job:

All commands below are for Claude Code. Other clients accept the same URLs and commands in their own config format.

How do you add an MCP server to Claude Code?

You add a server with claude mcp add, run from your terminal (not inside a session). Remote servers use --transport http and a URL; local servers use -- followed by the command that starts them (Claude Code MCP docs).

claude mcp add --transport http linear https://mcp.linear.app/mcpclaude mcp add playwright -- npx @playwright/mcp@latestclaude mcp list

After adding a server that uses OAuth, open a Claude Code session and run /mcp to sign in.

Choose a scope with --scope:

A committed .mcp.json is the cleanest way to give a team the same setup. Claude Code expands environment variables in it, so secrets stay out of git:

{  "mcpServers": {    "github": {      "type": "http",      "url": "https://api.githubcopilot.com/mcp/",      "headers": {        "Authorization": "Bearer ${GITHUB_PAT}",        "X-MCP-Toolsets": "repos,issues,pull_requests",        "X-MCP-Readonly": "true"      }    },    "context7": {      "type": "http",      "url": "https://mcp.context7.com/mcp"    },    "playwright": {      "command": "npx",      "args": ["@playwright/mcp@latest", "--headless", "--isolated"]    }  }}

Claude Code asks each teammate to approve project servers the first time it sees them.

Which MCP servers are best for code and repositories?

For repository work, the GitHub MCP Server is the most useful addition, because it lets the agent read issues, pull requests, Actions runs and code security alerts without you pasting them in. The Filesystem and Git reference servers matter mainly for clients that can't already read files and run git.

GitHub MCP Server

GitHub maintains its own server, with a hosted endpoint at https://api.githubcopilot.com/mcp/ (remote server docs). Two settings make it safer and lighter:

• X-MCP-Toolsets limits which groups of tools load, for example repos,issues. Fewer tools means less context used.
• X-MCP-Readonly: true enables only read tools. You can also append /readonly to the URL.

Use the read-only mode for agents that triage or summarize, and give write access only to agents whose job is to open pull requests.

Filesystem and Git

The MCP reference servers include Filesystem (file operations with configurable access controls) and Git (read, search and edit git repositories). They are useful in chat apps such as Claude Desktop that have no file access of their own. In a terminal agent they duplicate built-in tools, so skip them there.

Which MCP servers are best for docs and the web?

For documentation and web access, Context7 gives the agent current, version-specific library docs, and Playwright MCP lets it drive a real browser. Together they fix two of the most common agent failures: using an outdated API and declaring a UI change done without looking at it.

Context7

Context7 pulls up-to-date documentation and code examples for libraries straight from the source (Context7 repository). The remote server is at https://mcp.context7.com/mcp; an API key from the Context7 dashboard is recommended for higher rate limits. You can also run npx ctx7 setup, which signs you in and installs it for your agent.

claude mcp add --scope user --transport http context7 https://mcp.context7.com/mcp \  --header "Authorization: Bearer $CONTEXT7_API_KEY"

A prompt that uses it well names the library and version:

Use Context7 to look up the Next.js 16 docs for caching, then updateapp/products/page.tsx to match the current API. Cite the doc section you used.

Playwright MCP

Microsoft's Playwright MCP controls a browser through Playwright's accessibility tree instead of screenshots, so it works without a vision model (Playwright MCP). The agent can open your dev server, click through a flow, fill forms and report what broke. Add --headless on servers and --isolated to keep the browser profile in memory.

claude mcp add playwright -- npx @playwright/mcp@latest --headless --isolated

If you want agents to write and repair end-to-end tests with it, see AI test automation with Playwright.

Chrome DevTools MCP and Fetch

Chrome DevTools MCP (npx -y chrome-devtools-mcp@latest) adds performance traces, network requests and console messages with source-mapped stack traces, which makes it the better pick for debugging slow or broken pages. Its README warns that it exposes the browser's content to the MCP client, so don't use it in a browser profile signed into personal accounts. It also collects usage statistics by default; --no-usage-statistics turns that off.

Fetch is a reference server that retrieves a web page and converts it into text an LLM can read. Use it in clients without built-in web access.

Which MCP servers are best for databases and backends?

For data, the best choice is the official server for the platform you run on, connected in read-only mode by default. Supabase and Stripe both maintain hosted servers with scoping and approval controls built in.

Supabase MCP

Supabase hosts its server at https://mcp.supabase.com/mcp and supports two query parameters worth always setting (Supabase MCP docs): project_ref limits the server to one project, and read_only=true runs queries as a read-only Postgres user.

claude mcp add --scope project --transport http supabase \  "https://mcp.supabase.com/mcp?project_ref=YOUR_PROJECT_REF&read_only=true"

Supabase's own guidance is to connect to production only when the task needs production evidence, and to use a development branch for schema changes.

Stripe MCP

Stripe's server at https://mcp.stripe.com can search Stripe docs, read API data and, with permission, write it (Stripe MCP docs). Stripe requires human confirmation before some write actions such as refunds. As of October 2026, Stripe says the server will stop accepting full-access secret keys and untagged restricted keys after November 30, 2026; use OAuth or an agent API key.

claude mcp add --transport http stripe https://mcp.stripe.com/

What about plain PostgreSQL?

The original PostgreSQL reference server is archived. If you are not on a platform with an official server, connect a maintained community server with a read-only database user and point it at a staging copy, never at production with write access. For writing SQL with AI more generally, see our guide to an AI SQL query generator.

Which MCP servers are best for issues, errors and design?

The project-tool servers worth installing are the ones that save you copy-pasting: Sentry for errors, Linear (or your tracker) for tickets, Notion for specs and Figma for designs. All four are hosted by the vendor and use OAuth.

Sentry lets you append your organization and project slugs to the URL to scope the server (Sentry MCP). Figma's server needs a paid Figma plan, and Figma recommends the remote server over the desktop one for most users.

A Sentry-plus-GitHub prompt shows why these pay off:

Look up the top unresolved Sentry issue in project web-app from the last 24 hours.Find the code path in this repo, write a failing test that reproduces it,fix it, and open a draft PR that links the Sentry issue.

How many MCP servers should you install?

Install as few as you need, usually three to six. Each server adds tool names, descriptions and schemas to the agent's context, and large tool results can crowd out the code the agent is working on.

Claude Code has guards for this. Tool search is on by default, so the agent loads tool definitions on demand instead of all at once. Claude Code warns when one tool result passes 10,000 tokens, caps MCP output at 25,000 tokens by default (change it with MAX_MCP_OUTPUT_TOKENS), and saves very long text results to a file instead of the context (MCP docs).

Practical rules:

• Disable servers you haven't used this week. claude mcp remove <name> takes a second.
• Restrict toolsets where a server allows it, as GitHub does.
• Put personal servers in local or user scope and only shared, low-risk servers in .mcp.json.

Why context size matters so much is covered in our guide to context engineering.

How do you use MCP servers safely?

Treat every MCP server as code that runs with your permissions and every tool result as untrusted input. The main risks are an agent writing where it should only read, credentials leaking, and prompt injection through content a server fetches.

• Prefer official, vendor-hosted servers. They use OAuth, can be revoked from the vendor's dashboard and get security fixes.
• Use read-only modes first. GitHub (X-MCP-Readonly), Supabase (read_only=true) and read-only database users all limit the damage of a bad call.
• Scope to one project or org. Supabase's project_ref and Sentry's URL slugs keep an agent out of everything else.
• Keep secrets in environment variables, referenced as ${VAR} in .mcp.json, never pasted into committed files.
• Watch for injection. Claude Code's docs warn that servers fetching external content can expose you to prompt injection. A GitHub issue or web page can contain instructions aimed at your agent. Read our guide to prompt injection in coding agents before pairing a web-reading server with write access to anything.
• Review project servers before approving them. A cloned repo's .mcp.json can start local commands. Claude Code asks before using them; read the file before you say yes.

If you run agents unattended on a remote machine, the same rules apply with higher stakes, because nobody is watching each tool call. Hosted boxes such as Orkio isolate the agent's server from your laptop, but scoping each server's access is still your job.

Starter MCP stacks to set up today

Pick the stack that matches your work, add it in ten minutes, and use it for a week before adding more.

1. Web app developer: GitHub (read-only to start), Playwright, Context7, Sentry.
2. Backend or data developer: GitHub, Supabase or a read-only database server, Context7, Sentry.
3. Product engineer on a team: GitHub, Linear, Notion, Figma, committed to .mcp.json so everyone gets the same tools.
4. Payments work: add Stripe in a test environment first, with OAuth.

Then run this checklist:

• claude mcp list shows every server as connected.
• Secrets live in environment variables, not in .mcp.json.
• Write access is enabled only where an agent's task needs it.
• You have tried one real task per server and kept only the ones that saved time.

Frequently asked questions