Best MCP Servers for Coding on Mac
The best MCP servers for coding with Claude Code and Codex on a Mac, including Context7, Playwright and GitHub, with the command to add each one.
An MCP server is a small connector that gives an AI coding agent, such as Claude Code or Codex, an ability it doesn't have on its own. There are tens of thousands of them, and many are written by individuals rather than the companies whose products they connect to. Here are the ones I recommend for coding, with the commands to add them to Claude Code and Codex. If MCP is new to you, start with MCP Servers on Mac.
Adding MCP servers is one step to set up a Mac for software development. See the full roadmap for the rest.
Before you get started
I recommend using our free app to add your first MCP servers. It is a FREE Mac app that installs the developer tools and AI coding agents you need, adds MCP servers to Claude Code and Codex, and verifies your setup is working. Here are all the details about the app before you download. Then download it:
How I chose these MCP servers
Every server on this page is published by the company that makes the product. That is the best single sign that a server is safe to use and will be kept up to date.
To see which ones developers actually use, I checked two numbers in October 2026: how many people star each server's code on GitHub, and how many people search for it on Google each month in the US. I've added a few personal favorites as well, from lesser-known companies.
Don't add them all. Each MCP server will consume context (for an explanation, see MCP or Skills or CLI?). Add the ones that match your work, and remove the ones you stop using.
Most of the claude mcp add commands below add a server to the current project only. Add --scope user to make a server work in every project. A server you add to Codex works everywhere. After you add a server, restart Claude Code or Codex. If the server asks you to sign in, type /mcp in Claude Code and choose it, or run codex mcp login <name> for Codex. See Add MCP Servers to Claude Code and Claude Desktop for more.
More MCP servers
For directories of other MCP servers, and how to judge whether one is safe, see MCP Server Directories and Lists.
Context7 for library documentation
AI models learn from data that is months or years old, so they write code for old versions of libraries and frameworks. Context7, from Upstash, gives your agent the current documentation for the version you're using. It has the most GitHub stars of any server on this page, about 63,000. It works without a key, and a free key from context7.com raises its limits.
For Claude Code, then for Codex:
$ claude mcp add --transport http context7 https://mcp.context7.com/mcp
$ codex mcp add context7 --url https://mcp.context7.com/mcp
Upstash notes that the documentation comes from the community, so it can't promise every page is accurate.
Web fetch and search
AI agents can fetch web pages on their own, but they often get raw page code, a page that blocks them, or nothing at all. A web-reading MCP server gives your agent clean text.
Firecrawl to read web pages
Firecrawl reads pages, including ones built with JavaScript, and gives your agent clean Markdown. Its remote server works without a key for reading a page, searching, and reading files such as PDFs, at a limited rate. A free account (1,000 credits a month, no credit card) unlocks its full set of tools.
For Claude Code, then for Codex, without a key:
$ claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp
$ codex mcp add firecrawl --url https://mcp.firecrawl.dev/v2/mcp
To use your account, add the server with the address that ends in mcp-oauth and sign in. Firecrawl warns never to put an API key in the server's web address or in a chat with your agent.
Jina Reader to read web pages and search
Jina Reader, from Jina AI, reads web pages and PDFs as Markdown and searches the web. It needs a free API key from jina.ai. When I checked in October 2026, its page-reading and search tools refused every request without one, although its instructions say they work without a key.
For Claude Code:
$ claude mcp add --scope user --transport http jina https://mcp.jina.ai/v1 --header "Authorization: Bearer YOUR_API_KEY"
For Codex, add these lines to ~/.codex/config.toml:
[mcp_servers.jina]
url = "https://mcp.jina.ai/v1"
http_headers = { "Authorization" = "Bearer YOUR_API_KEY" }
Brave Search for web search
Brave Search gives your agent web search results from Brave's own index. It runs on your Mac, so it needs Node.js (see the note at the end of this page), and it needs an API key from Brave. The free credits cover about 1,000 searches a month, but Brave asks for a credit card to confirm who you are. Brave doesn't give the commands for Claude Code or Codex, so these are mine, following each tool's rules:
$ claude mcp add --env BRAVE_API_KEY=YOUR_API_KEY --transport stdio brave-search -- npx -y @brave/brave-search-mcp-server
$ codex mcp add brave-search --env BRAVE_API_KEY=YOUR_API_KEY -- npx -y @brave/brave-search-mcp-server
Test and control a web browser
Playwright to test your web app
Playwright MCP, from Microsoft, lets your agent open a web page, click through it, fill in forms, and see what happens, which is how it tests the web app it just built. It runs on your Mac with Node.js 18 or later, and it uses your installed Google Chrome unless you choose another browser.
For Claude Code, then for Codex:
$ claude mcp add playwright npx @playwright/mcp@latest
$ codex mcp add playwright -- npx @playwright/mcp@latest
Microsoft says plainly that Playwright MCP is not a security boundary. It keeps the browser's logins between sessions, so don't sign in to accounts you don't want your agent to use. Add --isolated after @playwright/mcp@latest to start fresh each time. Microsoft also notes that coding agents can use the Playwright command-line tool with fewer tokens. See MCP or Skills or CLI.
Chrome DevTools to debug web pages
Chrome DevTools MCP, from Google's Chrome team, gives your agent the same developer tools you use in Chrome: console errors, network requests, screenshots, and performance traces. It is the tool I'd add when the agent needs to find out why a page is broken. It needs Google Chrome and Node.js 20 or later.
For Claude Code, then for Codex:
$ claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest
$ codex mcp add chrome-devtools -- npx chrome-devtools-mcp@latest
Google collects usage statistics from it by default. Add --no-usage-statistics after chrome-devtools-mcp@latest to turn that off.
GitHub's MCP server
Most software projects use GitHub for collaboration.
Many developers skip this MCP server and let the agent use the gh command-line tool instead, which the agent already knows. If you prefer to use an MCP server, GitHub's own MCP server lets your agent read and manage issues, pull requests, GitHub Actions runs, and code. From Claude Code and Codex, it signs in with a personal access token, which you create at GitHub's token settings. Give the token only the permissions you need.
For Claude Code, GitHub gives this command:
$ claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_TOKEN"}}'
For Codex, add these lines to ~/.codex/config.toml:
[mcp_servers.github]
url = "https://api.githubcopilot.com/mcp/"
http_headers = { "Authorization" = "Bearer YOUR_GITHUB_TOKEN" }
To start safely, use the read-only address, which ends in /mcp/readonly.
Connect your team's tools
These MCP servers connect your AI agent to tools used by workplace teams. Each one runs remotely and requires your own sign-in, so your agent sees only what you can see. Your company may need to allow AI tools before you can connect.
Notion's MCP server for docs and wikis
I think Notion is clunky and slow, but I've worked in organizations where Google Docs (or Markdown files on GitHub) are not an option for company-wide document sharing.
Notion's server lets your agent read your specs and notes in Notion, and write pages and database entries back. Notion warns that it can reach everything you can reach in Notion.
$ claude mcp add --transport http notion https://mcp.notion.com/mcp
$ codex mcp add notion --url https://mcp.notion.com/mcp
Linear's MCP server for issues and projects
I prefer Linear for issue tracking on a team, compared to Jira.
Linear's server lets your agent find, create, and update issues and projects. To let the agent read but never change anything, use the address that ends in /mcp/readonly.
$ claude mcp add --transport http linear https://mcp.linear.app/mcp
$ codex mcp add linear --url https://mcp.linear.app/mcp
Atlassian's MCP server for issues and documents
Like many developers, I really dislike Jira, Atlassian's issue tracking app. But Jira is more tolerable when you interact with an agent instead of the Jira UI.
Atlassian's server, called Rovo MCP, lets your agent read and update Jira work items and Confluence pages. It works with Atlassian's cloud sites only, and each tool your agent uses counts against your organization's Rovo credits. Atlassian keeps the tools that delete things turned off unless an administrator turns them on.
$ claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp
$ codex mcp add atlassian --url https://mcp.atlassian.com/v2/mcp
Slack's MCP server for team messages
Slack's server lets your agent search messages and threads for context, and post messages. It needs your workspace administrator to approve it. For Claude Code, Slack recommends its plugin, which handles the sign-in. Type this inside Claude Code:
/plugin install slack@claude-plugins-official
For Codex, Slack's instructions require a Slack app of your own. See Slack's guide to connecting AI tools.
Connect databases and app services
These servers can change real data, and one can move real money, so set them up with care. Approve each action the agent wants to take, and start with test data, not your live data.
Supabase's MCP server for a database
Supabase's server lets your agent read your database's structure, run SQL, manage migrations, and read logs. Supabase lets you limit it in the server's address. I recommend read_only=true, which runs every query as a read-only database user, and project_ref= with your project's ID, which limits the server to one project:
$ claude mcp add --transport http supabase "https://mcp.supabase.com/mcp?project_ref=YOUR_PROJECT_ID&read_only=true"
$ codex mcp add supabase --url "https://mcp.supabase.com/mcp?project_ref=YOUR_PROJECT_ID&read_only=true"
Supabase advises connecting to your live database only when a task needs live data.
InsForge's MCP server for app backends
InsForge is a backend as a service: a Postgres database, sign-in for your app's users, file storage, and server functions, all in one hosted project. It is built for AI coding agents, and it is open source, so you can also host it yourself. I use it myself for my own projects. Its MCP server lets your agent set up and manage your project's database, sign-in, storage, and functions. You sign in with your InsForge account in your browser, so there is no API key to copy.
For Claude Code, then for Codex:
$ claude mcp add --transport http insforge https://mcp.insforge.dev/mcp
$ codex mcp add insforge --url https://mcp.insforge.dev/mcp
InsForge also offers a command-line tool, npx @insforge/cli, which installs InsForge's agent skills for Claude Code and Codex.
Sentry's MCP server for error tracking
Sentry's server gives your agent an error's stack trace and details from Sentry, so it can find and fix the bug. It signs in with your Sentry account.
$ claude mcp add --transport http sentry https://mcp.sentry.dev/mcp
$ codex mcp add sentry --url https://mcp.sentry.dev/mcp
Stripe's MCP server for payment services
Stripe's server helps your agent build a Stripe integration: it looks up Stripe's API and documentation, and reads and creates objects in your Stripe account. When you sign in, Stripe asks which accounts to allow. Choose a sandbox (Stripe's test mode), not your live account, while you build. Stripe asks you to turn on your agent's approval prompts, and it requires your confirmation for actions such as refunds.
$ claude mcp add --transport http stripe https://mcp.stripe.com/
$ codex mcp add stripe --url https://mcp.stripe.com
Stripe is about real money and customers, so be careful when you use an AI agent.
Design tools
The product design profession is changing with AI, just as software engineering has changed. If you're a product designer, you may already be using Claude Design, Google Stitch, or another AI-first design tool. If your team uses Figma for interface and product design, or Canva for general graphic design, you can integrate either with your AI agent.
Figma's MCP server for product design
Figma's server gives your agent a design's layout, colors, and components, so it can build a matching interface. The remote server works on every Figma plan, but Figma limits how often View and Collab seats can use it. For Claude Code, Figma recommends its plugin. Here is the plugin command, then the command for Codex:
$ claude plugin install figma@claude-plugins-official
$ codex mcp add figma --url https://mcp.figma.com/mcp
Canva's MCP server for graphic design
Canva's server, which Canva calls its AI Connector, lets your agent search, create, and edit your Canva designs, fill in brand templates, and export files such as PDF, PNG, and PowerPoint. It works on every Canva plan, including the free one, but brand templates need a paid plan, and some actions use your monthly Canva AI credits. On a Canva team, an administrator has to turn it on. It signs in with your Canva account.
$ claude mcp add --transport http canva https://mcp.canva.com/mcp
$ codex mcp add canva --url https://mcp.canva.com/mcp
Install Node.js for local MCP servers
Brave Search, Playwright, and Chrome DevTools run on your Mac and start with the npx command, which comes with Node.js. Install the current version of Node.js with Homebrew:
$ brew install node
The other servers on this page run in the cloud and need nothing installed.
Continue setting up your Mac
Don't miss the full visual roadmap and checklist that shows how to set up a Mac for software development, with all the essential tools and settings you might not yet know about.