The best MCP servers for getting real work done in 2026

The best MCP servers in 2026 are the ones that touch your daily work: GitHub for code, Filesystem and Git for local projects, Playwright for the browser, Context7 for current library docs, Supabase or another database server for data, and Sentry for production errors. Connect three to five at a time, not thirty.

I run a handful of MCP servers every day, and most lists pad themselves to thirty entries nobody installs. This is the short version: what each server is good for, who maintains it (which changed a lot in the last year), and how to add one. heydecks, the deck tool I build, is on the list near the end because it does one narrow job.

#Which MCP servers are worth installing first?

Here is the table I would hand a teammate setting up Claude Code or Cursor. Star counts are from GitHub in September 2026.

None of these is the best in the abstract. The best one closes the gap between what your agent can see and what your job needs it to touch.

#Who maintains MCP servers now?

This changed, and many lists are out of date. The official modelcontextprotocol/servers repository now keeps only seven reference servers: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, and Time. The older reference servers for GitHub, PostgreSQL, Puppeteer, Brave Search, Sentry, Slack, SQLite, Google Drive, and others were moved to an archive repository.

Most of those jobs now have official servers built by the vendor: GitHub publishes github/github-mcp-server, Brave publishes its own search server, Sentry runs a hosted one, and Supabase ships one for its databases. So "the reference Postgres server" is no longer a safe recommendation; use your database vendor's server or a maintained community one. To find servers, start with the official MCP Registry at registry.modelcontextprotocol.io, then check who publishes a server before you give it write access.

#What do these servers actually do?

Code and repositories. GitHub is the most-installed server for a reason: the agent reviews issues, opens pull requests, and searches across repos without you pasting URLs. The hosted endpoint supports OAuth, so there is nothing to build. Filesystem and Git are the local pair: files in folders you allow, plus commits, diffs, and branch state.

Browser and web. Playwright gives the agent a real browser. It works from the page's accessibility tree rather than screenshots, so it can fill forms and run end-to-end tests without a vision model. Fetch is the light option: a URL in, markdown out. Brave Search adds live results for research that needs today's web.

Docs, data, and errors. Context7 is the one I would call underrated. It pulls current documentation for the library version you use, which stops the agent calling a method removed two releases ago. Supabase lets an agent read schemas and run queries; turn on read-only mode before pointing it at production. Sentry feeds real stack traces into the chat, so debugging starts from the actual error.

#What are the best MCP servers for coding?

GitHub, Filesystem, Git, Context7, and your database server cover most coding work. Add Playwright when you test UI and Sentry when you chase a production bug.

The part most lists skip: every connected server loads its tool descriptions into the model's context before you ask anything. Twenty servers quietly eat the budget the model needs for your code, and in my experience the agent gets slower and picks the wrong tool more often past a handful. Connect what the current task needs and switch the rest off. Claude Code and Cursor both let you toggle servers per project.

#Are MCP servers free and safe?

Most are free and open source. What you pay for is the service behind them: a database host, a Brave Search API key, a Sentry plan. heydecks works the same way: connecting is free, and building a deck spends credits (100 per deck), the same as in the dashboard or over the API.

Safety depends on the code and the access you grant. A server with write access to your repo or database is software you are trusting. Prefer vendor-published servers, check the scopes they ask for, use OAuth or short-lived tokens, and remember that any server returning text can try to steer the model. The what is MCP primer covers the trust model.

#How do you add an MCP server to Claude, Cursor, or VS Code?

Usually one command or a few lines of config. In Claude Code, a remote server is one line:

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

The first call opens a browser to sign in. Cursor takes the URL as JSON and runs the sign-in for you:

{ "mcpServers": { "heydecks": { "url": "https://heydecks.com/mcp" } } }

In claude.ai and Claude Desktop, add the URL as a custom connector under Customize, Connectors. Local servers use a command and args entry instead of a URL. Whatever the host, restart it, confirm the server appears, and start with a read-only request before anything destructive. Per-app details are in Claude MCP setup.

#Can an MCP server build a slide deck?

Yes, and it is the one job I can speak to directly. heydecks is the AI slide creator for AI agents: the agent writes the content and picks from 95 hand-designed slide templates, and heydecks renders the deck on your brand via the Brand Kernel, returning a live link, a PDF, and (on Solo and up) a native, editable .pptx. You reach for it at the end of a task, once the analysis is done and what is left is slides someone can present.

An agent that just pulled this week's numbers from Supabase and Sentry can hand them to heydecks and get back a weekly report like the one below (built from the same templates, on a sample brand), instead of you exporting a CSV and fighting a template. PowerPoint MCP servers compares the other deck options, and building decks with Claude has prompts to start from.

If a deck is the last step of your agent's work, the heydecks MCP server page has the URL and the full tool list.

#Frequently asked questions

#What is the most popular MCP server?

By GitHub stars, Context7 (about 62,500) and Playwright (about 37,700) lead, with GitHub's official server close behind at about 33,000. GitHub is still the one most developers connect first, because so much agent work happens in repos and pull requests.

#How many MCP servers should I install?

Three to five active at a time for most setups. Each server's tool descriptions load into the model's context up front, so a long list costs you context and tool-picking accuracy. Keep others installed but switched off.

#Where can I find MCP servers?

The official MCP Registry at registry.modelcontextprotocol.io is the best starting point, since it is run by the MCP project. The modelcontextprotocol/servers repository lists the seven reference servers and links out to vendor servers.

#Is there still an official Postgres MCP server?

Not from the MCP project. The reference PostgreSQL server was archived along with GitHub, Puppeteer, Brave Search, and others. Use your database vendor's server, such as Supabase's, or a maintained community server you have vetted.

#Can an MCP server make a PowerPoint?

Yes. heydecks runs a hosted MCP server that builds a branded deck from your agent's content and exports a PDF and a native, editable .pptx. Open-source python-pptx servers also write .pptx files if you style them yourself.