The best MCP servers to install first are the official ones for tools you already use: GitHub for code, Playwright for the browser, Context7 for current library docs, and the servers from Sentry, Linear, Figma or your database provider. Start with two or three that match your daily work, not twenty. Every server adds tools the model has to read, and access you have to trust.
Every server below is published by the company behind the product or maintained by the MCP project itself. We checked each one against its own repository or docs in September 2026.
- Pick servers by job, and start with two or three. More tools means more for the model to read and more access to trust.
- Prefer servers from the product’s own team. The MCP project’s reference servers are teaching examples, not products.
- For online services, prefer remote servers with OAuth sign-in, and use read-only modes when the agent only needs to look.
- Never give one agent private data, untrusted content and a way to send data out, all at once.
- Vet every server before you add it: publisher, activity, exact command, permissions and what leaves your machine.
The shortlist at a glance
If MCP is new to you, start with what the Model Context Protocol is. Here is the whole list on one screen.
| Job | Server | Published by | Runs |
|---|---|---|---|
| Files | Filesystem | MCP project (reference) | Locally |
| Code and pull requests | GitHub MCP Server | GitHub | Remote or local |
| Browser automation | Playwright MCP | Microsoft | Locally |
| Browser debugging | Chrome DevTools MCP | Google’s Chrome DevTools team | Locally |
| Library docs | Context7 | Upstash | Remote |
| Supabase projects | Supabase MCP | Supabase | Remote |
| Other databases | MCP Toolbox for Databases | Locally | |
| Design files | Figma MCP server | Figma | Remote |
| Issue tracking | Linear MCP | Linear | Remote |
| Jira and Confluence | Atlassian MCP Server | Atlassian | Remote |
| Error monitoring | Sentry MCP | Sentry | Remote |
To add a remote server, paste its URL into your app’s MCP settings. In Claude Code, for example, claude mcp add --transport http linear https://mcp.linear.app/mcp adds Linear, and running /mcp in a session signs you in.
Files, Git and GitHub
Reads, writes, moves and searches files, but only inside the folders you list in its config. It is a reference server from the MCP project, which calls these teaching examples rather than production-ready products. Fine for a notes folder or a sandbox. Keep the folder list short.
GitHub’s official server for repositories, issues, pull requests, Actions runs and security alerts. Connect to https://api.githubcopilot.com/mcp/, or run it locally. Turn on only the toolsets you need, which GitHub says also helps the model choose tools. Use read-only mode when the agent only needs to look.
For local Git history without GitHub, the MCP project’s reference Git server (uvx mcp-server-git) can read, search and change repositories. The same teaching-example caveat applies.
Browser automation and debugging
Microsoft’s server lets an agent drive a real browser: open pages, click, type and fill in forms. It works from the page’s accessibility tree rather than screenshots, which makes it fast and precise. Start it with npx @playwright/mcp@latest.
Lets an agent inspect a live Chrome tab: console messages, network requests, screenshots and performance traces. It can see and change everything in that browser, so keep personal accounts out of it. Usage statistics go to Google by default. Add --no-usage-statistics to opt out.
Microsoft notes that coding agents increasingly favor Playwright’s command-line tool plus a skill, because it loads less into the context window. If your agent supports them, see how agent skills work.
Up-to-date docs and databases
Pulls current, version-specific documentation and code examples into the prompt, so the agent stops guessing at APIs that changed after its training. Add “use context7” to a request, or set a rule to use it by default. The endpoint is https://mcp.context7.com/mcp, and a free API key raises the rate limits.
Supabase’s official server lets an agent manage tables, fetch config and query data in your projects. URL options limit it to one project and make it read-only, such as ?project_ref=<ref>&read_only=true. Supabase recommends reading its security guidance before you connect.
Google’s open-source server connects agents to PostgreSQL, MySQL, SQL Server, MongoDB, Snowflake, BigQuery and many more. One flag, such as --prebuilt=postgres, gives an agent ready-made tools to explore schemas and query data. It also works as a framework for building narrower tools for production agents.
Design, issues and errors
Figma’s official server gives coding agents the real design context of a frame, including variables, components and layout, so generated code matches the design. The endpoint is https://mcp.figma.com/mcp. Check your seat first: Starter plans and View or Collab seats get up to six tool calls a month.
Linear’s hosted server finds, creates and updates issues, projects and comments. Connect to https://mcp.linear.app/mcp and sign in with OAuth. When the agent should only read, use https://mcp.linear.app/mcp/readonly instead.
Atlassian’s official cloud server covers Jira, Confluence, Jira Service Management, Bitbucket, Compass and Loom. Use the v2 endpoint, https://mcp.atlassian.com/v2/mcp, which Atlassian now recommends. Actions respect your existing Atlassian permissions.
Sentry’s hosted server lets a coding agent look up errors, issues and traces from production, then fix the code next to them. Connect to https://mcp.sentry.dev/mcp. Sentry built it for human-in-the-loop coding agents, so keep a person reviewing the fix.
How to judge an MCP server before you install it
This list will age. Here is how to vet anything that is not on it.
Check who publishes it
Prefer the product’s own organization, such as
github/,getsentry/oratlassian/. In the official MCP Registry, a name likeio.github.user/serveris tied to a verified account or domain. That proves who published it, not that the code is safe.Check that it is alive
Look for recent commits and releases, and issues that get answers. The MCP project has archived its early GitHub, Postgres, Puppeteer and Sentry servers, so old tutorials often point at dead code.
Prefer remote for online services
A hosted server with OAuth sign-in runs none of its code on your machine, and you can revoke access from the service’s settings. Keep local servers for local jobs, such as files and browsers.
Read the exact command or URL
Copy it from the official README, not a blog or a forum. Lookalike packages exist, so check the package name letter by letter.
Count the tools
In most apps, every tool description sits in the model’s context. Fewer tools means a cheaper, sharper agent, so switch off toolsets you do not use.
Check what leaves your machine
Read the README’s privacy notes. Some servers send usage statistics or call outside services unless you opt out.
If none of these cover your job, building your own is not hard. Our tutorial on building an MCP server walks through a working one in about 30 minutes.
The permissions to watch
Simon Willison calls the worst combination the lethal trifecta: an agent with access to your private data, exposure to untrusted content, and a way to send data out. MCP makes this easy to assemble by accident, because each server adds one piece.
Picture a GitHub server that can read your private repositories, also reads public issues that anyone can write, and can post comments. A hidden instruction in one issue could tell the agent to copy private code into a public comment. That is prompt injection, and no single server is to blame.
FAQ
How many MCP servers should I install?
As few as your work needs, often two or three. Each server adds tool descriptions to the model’s context, and GitHub notes that limiting toolsets also helps the model choose the right tool.
Are MCP servers free?
Most servers are free and many are open source, but the service behind them may need a paid plan. Figma, for example, limits tool calls on its Starter plan and for View or Collab seats.
What happened to the GitHub and Postgres servers in the MCP repo?
They were archived. GitHub now maintains its own official server. For browsers, use Playwright MCP or Chrome DevTools MCP. For Postgres, use your database provider’s server or MCP Toolbox for Databases.
Where can I find more servers?
The official MCP Registry lists public servers and verifies who published each one. It is still in preview and does not scan code, so vet anything you find there.
Do these servers work in every AI app?
They work in any app that supports MCP, though each app has its own settings screen or config file. Remote servers need an app that supports remote connections, which most major ones now do.
- Start with two or three official servers that match your daily work.
- Prefer hosted servers with OAuth for online services, and local ones for files and browsers.
- Use read-only modes, project scoping and small toolsets.
- Vet the publisher, activity, command and data flows before you install.
- Never combine private data, untrusted content and an outbound channel in one agent.
Read next: what MCP is and how it works, or why prompt injection is the risk to understand before you connect anything.
- Model Context Protocol servers, MCP project, September 2026
- GitHub MCP Server, GitHub, September 2026
- Playwright MCP, Microsoft, September 2026
- Chrome DevTools MCP, Google, September 2026
- Context7, Upstash, September 2026
- Supabase MCP server, Supabase, September 2026
- MCP Toolbox for Databases, Google, September 2026
- Figma MCP server guide, Figma, September 2026
- MCP server, Linear
- Atlassian MCP Server, Atlassian, September 2026
- Sentry MCP, Sentry, September 2026
- The MCP Registry, Model Context Protocol
- The lethal trifecta for AI agents, Simon Willison, June 2025
