12 Best MCP Servers for Coding and Building in 2026

12 Best MCP Servers for Coding and Building in 2026

Short answer

The MCP servers most worth installing for coding are the GitHub MCP Server (issues, pull requests, Actions), Playwright MCP or Chrome DevTools MCP (a real browser for testing and debugging), and Context7 (current library docs). Add the official server for each service you already run: Sentry, Supabase, Figma, Linear, Stripe or Vercel. Every server below is maintained by its vendor or by the MCP project, and each was checked on its official repo or docs in November 2026.

A coding agent is only as useful as what it can reach. With the right three or four servers, it can read the failing issue, reproduce the bug in a browser, check the Sentry trace and open the pull request. With the wrong twenty, it burns context on tool lists and holds tokens it never needed.

What an MCP server does

The Model Context Protocol is an open standard for connecting AI apps to tools and data. A server exposes tools (actions the model can call), resources (data it can read) and prompts. A client, such as Claude Code, Cursor or Copilot in VS Code, connects to the server and offers those tools to the model. If the protocol is new to you, start with our explainer on what MCP is and why it matters for app builders.

Servers come in two shapes. A local server runs on your machine as a process the client starts (a command like npx or uvx) and talks over stdio. A remote server runs at a URL over Streamable HTTP and usually signs you in with OAuth, so there is no token to paste.

How we picked these servers

Each server had to help with writing, testing, shipping or debugging code, and be maintained by its vendor or by the MCP project's steering group. We read each one's official repo or docs in November 2026 for maintainer, tools, install path, sign-in and limits, and left out star counts.

We also dropped anything archived. The MCP project has moved its original GitHub, GitLab, PostgreSQL, Puppeteer, Redis, Sentry, Slack and SQLite reference servers into an archive of servers that are no longer maintained, so older guides that recommend them are out of date. Stripe has also deprecated its local @stripe/mcp package in favor of its hosted server. Community servers exist for almost everything and the official MCP Registry lists published ones, but a vendor-run server is the safer default when one exists.

At-a-glance comparison

ServerMaintained byWhat it gives your agentHow it connects
GitHub MCP ServerGitHub (official)Repos, issues, pull requests, Actions, security alertsRemote with OAuth or a PAT, or local in Docker
Playwright MCPMicrosoft (official)Browser control through accessibility snapshotsLocal, npx @playwright/mcp@latest
Chrome DevTools MCPChrome DevTools project (official)Performance traces, network, console, screenshotsLocal, npx chrome-devtools-mcp@latest
Context7Upstash (official)Current, version-specific library docsRemote with an API key, or npx ctx7 setup
FilesystemMCP steering group (reference)Read, write and search files in allowed foldersLocal, npx @modelcontextprotocol/server-filesystem
GitMCP steering group (reference)Status, diffs, log, commits, branchesLocal, uvx mcp-server-git
Sentry MCPSentry (official)Errors, traces, issue triageRemote with OAuth, or local through the Sentry CLI
Supabase MCPSupabase (official)SQL, migrations, logs, advisors, Edge FunctionsRemote with OAuth, or a PAT in CI
Figma MCP serverFigma (official)Design context for code; writes to the canvasRemote, or desktop app on paid seats
Linear MCPLinear (official)Find, create and update issues and projectsRemote with OAuth 2.1 or an API key
Stripe MCPStripe (official, public preview)Stripe API reads and writes, docs searchRemote with OAuth or an agent API key
Vercel MCPVercel (official)Docs, projects, deployments, logs, analyticsRemote with OAuth

The best MCP servers for coding

1. GitHub MCP Server: your repos, issues and pull requests

GitHub's official server lets an agent browse code, manage issues and pull requests, check Actions runs and read security findings such as Dependabot alerts. The hosted version lives at api.githubcopilot.com/mcp/ and signs in with OAuth or a personal access token; a local version runs from the ghcr.io/github/github-mcp-server Docker image. With no options set, it loads the context, repos, issues, pull_requests and users toolsets.

Limits: OAuth support depends on your client, since each host has to set up its own GitHub or OAuth app. A token gives the server whatever you grant it, so trim permissions, load only the toolsets you need with --toolsets, and use --read-only when the agent only needs to look.

2. Playwright MCP: a browser the agent can drive

Microsoft's Playwright MCP controls a real browser through Playwright's accessibility tree rather than screenshots, so the model needs no vision support. It needs Node.js 18 or newer and installs with one line in most clients, for example claude mcp add playwright npx @playwright/mcp@latest.

Limits: Microsoft's README says the server is not a security boundary, and that coding agents may do better with its Playwright CLI and skills, which use fewer tokens than MCP tool schemas and accessibility trees. A saved browser profile works for one client at a time; use --isolated for parallel runs.

3. Chrome DevTools MCP: debugging and performance

Published by the Chrome DevTools project, this server gives an agent a live Chrome with DevTools: it records performance traces, inspects network requests, reads console messages with source-mapped stack traces and takes screenshots. Add it with npx -y chrome-devtools-mcp@latest; it needs a Node.js LTS version and current Chrome.

Limits: Google collects usage statistics by default (turn off with --no-usage-statistics), and performance tools may send trace URLs to the Chrome UX Report API (--no-performance-crux). It officially supports Chrome and Chrome for Testing only, and the browser's contents are visible to your MCP client.

Build this with a prompt on Greta.sh — describe the app, Greta writes and deploys it.

4. Context7: current docs instead of stale training data

Context7, from Upstash, pulls version-specific documentation and code examples into the agent's context through two tools, resolve-library-id and query-docs. Connect to mcp.context7.com/mcp with an API key in the Authorization header, or run npx ctx7 setup, which signs you in and configures Cursor, Claude Code or OpenCode. A free API key raises your rate limits.

Limits: library entries are community-contributed, and Upstash says it can't guarantee their accuracy, completeness or security. The repo holds the MCP server only; the API backend and crawler are private.

5. Filesystem: files in folders you choose

A reference server from the MCP project for reading, writing, moving and searching files. It touches only the folders you pass on the command line, and --read-only (or --read-only-dir per folder) withholds or blocks the write tools. Most useful in clients that can't edit files themselves, such as chat apps.

Limits: the MCP project calls its reference servers educational examples, not production-ready solutions, and asks you to apply your own safeguards.

6. Git: local history and diffs

The Git reference server exposes git_status, git_diff, git_log, git_show, git_commit, branch tools and two prompts for commit messages and change summaries. Run it with uvx mcp-server-git --repository [path].

Limits: it is in early development, and without --repository it can read and change any repository the process can reach. It never runs hooks or filter programs from a repository's own config.

7. Sentry MCP: errors and traces in your editor

Sentry built its remote server for human-in-the-loop coding agents: search errors, analyze performance, triage issues and read Sentry docs. Connect to mcp.sentry.dev/mcp with OAuth, ideally scoped to one organization and project, which Sentry recommends. Self-hosted Sentry users can run it locally with the Sentry CLI's sentry mcp command.

Limits: it is not a general-purpose Sentry API, and when you run it locally its AI-powered search tools need an LLM provider key (OpenAI, Azure OpenAI, Anthropic or OpenRouter).

8. Supabase MCP: your database and backend

Supabase's server, now at github.com/supabase/mcp, runs at mcp.supabase.com/mcp and asks you to log in when you connect. Its tools run SQL, list and apply migrations, query logs, fetch security and performance advisors, generate TypeScript types and manage Edge Functions. Branching tools need a paid plan.

Limits: Supabase says production projects can hold sensitive data: scope the server to one project, turn on read-only mode and restrict feature groups before connecting one. The CLI and self-hosted versions offer fewer tools and no OAuth 2.1.

9. Figma MCP server: designs to code and back

Figma's server gives agents design context from selected frames, variables and components, and can now write native content back to the canvas. The remote server at mcp.figma.com/mcp is Figma's recommended option; the desktop server needs the Figma desktop app and a Dev or Full seat on a Professional, Education, Organization, Government or Enterprise plan.

Limits: read tools are rate limited. The Starter plan and View or Collab seats get up to 20 tool calls a month; Dev and Full seats get up to 200 a day on Professional and Organization plans and up to 600 on Enterprise. Writing to the canvas is in beta, free for now, and Figma says it will become a usage-based paid feature.

10. Linear MCP: issues and projects

Linear hosts its server at mcp.linear.app/mcp over Streamable HTTP. It signs in with OAuth 2.1, or takes a Linear API key or OAuth token as a bearer header, and its tools find, create and update issues, projects and comments. In Claude Code: claude mcp add --transport http linear-server https://mcp.linear.app/mcp.

Limits: for look-only access, connect to /mcp/readonly or grant only the read scope. The old /sse endpoint is deprecated, and clients without remote support need the mcp-remote bridge.

Want to try one of these? Start on Greta.sh — no setup, no boilerplate — just a prompt.

11. Stripe MCP: payments with human confirmation

Stripe's hosted server at mcp.stripe.com is in public preview. It can search, read and write the Stripe API and search Stripe's docs, and some actions, such as refunds, wait for you to approve them in the Dashboard. Use OAuth for interactive clients: it lets you authorize and revoke a client without sharing an API key.

Limits: from December 2026 it stops accepting full-access secret keys and restricted keys without the Agent tag; Stripe asks users to switch to agent API keys or OAuth by November 30, 2026. Stripe also tells you to keep tool approval on and to watch for prompt injection when it runs next to other servers.

12. Vercel MCP: deployments and logs

Vercel's official server at mcp.vercel.com uses OAuth and lets an agent search Vercel docs, manage projects and deployments, analyze deployment logs and query Web Analytics. Vercel says it is available on all plans, subject to its AI Product Terms.

Limits: it accepts only clients Vercel has reviewed (Claude Code, Cursor and VS Code with Copilot are on its list) or clients that identify themselves with a metadata document. Vercel also notes that connecting gives the AI tool the same access as your Vercel account.

How to add an MCP server

The three clients most developers use each take a URL for remote servers and a command for local ones. These examples follow each client's own MCP docs as of November 2026.

Claude Code

Add a remote server with --transport http, or a local one after a double dash. Each command writes to local scope unless you add --scope project (a shared .mcp.json) or --scope user. Inside a session, /mcp shows server status and handles sign-in.

# remote server, shared with your team through .mcp.json
claude mcp add --transport http --scope project linear https://mcp.linear.app/mcp

# local server: everything after -- is the launch command
claude mcp add playwright -- npx @playwright/mcp@latest

# check what is connected
claude mcp list

Cursor

Cursor reads ~/.cursor/mcp.json for servers you want everywhere and .cursor/mcp.json in a project for project-only servers. You can also install from its marketplace with Add to Cursor. Use ${env:NAME} instead of pasting a key into the file.

{
  "mcpServers": {
    "sentry": {
      "url": "https://mcp.sentry.dev/mcp/[org-slug]/[project-slug]"
    },
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest"]
    },
    "context7": {
      "url": "https://mcp.context7.com/mcp",
      "headers": { "Authorization": "Bearer ${env:CONTEXT7_API_KEY}" }
    }
  }
}

VS Code (GitHub Copilot)

Run MCP: Add Server from the Command Palette and enter a URL or a launch command. Choose Global, then Copilot Global, for a personal setup across your workspaces, or a workspace-root .mcp.json to share a server with teammates.

Command Palette (Cmd+Shift+P / Ctrl+Shift+P)
> MCP: Add Server
  1. Choose a URL (HTTP) or a command (stdio)
  2. Enter https://api.githubcopilot.com/mcp/ or npx -y @playwright/mcp@latest
  3. Enter a server ID, such as github
  4. Pick where to save it: Global > Copilot Global, or the workspace .mcp.json

Security: what to check before you connect one

The MCP project's security guidance is blunt about local servers: they run on your machine and, without sandboxing and consent, a malicious launch command can run any command with the client's privileges. VS Code's docs say the same in one line: local servers can run code on your machine, and tools act with your permissions.

  • Install from the source. Copy the URL or package name from the vendor's repo or docs, not from a forum post, and double-check the domain on one-click installs.
  • Grant the least access. Use read-only modes where they exist (GitHub, Supabase, Linear, Filesystem) and tokens scoped to one project.
  • Keep approvals on. Cursor asks before running MCP tools by default, and Claude Code asks you to approve servers from a project's .mcp.json; a cloned repo can't approve its own.
  • Watch for prompt injection. A server that reads issues, web pages or logs can bring in text written by someone else. Claude Code's docs warn about servers that fetch outside content for this reason.
  • Keep secrets out of shared config. Use environment variables (${env:NAME} in Cursor) and never commit a token in .mcp.json.
  • Prune what you don't use. Every connected server adds tools the model reads, and access you have to keep track of.

How to choose

Start with the work you repeat. Hand-copying issue text into prompts? Add GitHub or Linear. Checking UI changes by hand? Add Playwright, or Chrome DevTools for network and performance problems. Agent writing code for an old library version? Add Context7. Then add the server for the one service you debug most, and stop until you feel a gap.

Already working with an agent? Our Claude Code tips cover MCP scopes and permission rules, and our AI coding assistants roundup compares the clients themselves. For the habits that keep agent output reviewable, read our AI coding best practices.

Frequently Asked Questions

What is an MCP server?
An MCP server is a program that exposes tools, data or prompts to an AI app through the Model Context Protocol. A coding agent such as Claude Code, Cursor or Copilot in VS Code connects to it and can then call its tools, for example to read a GitHub issue, query a Supabase database or open a page in a browser.

Are MCP servers safe to use?
They can be, with care. A local server runs code on your machine with your permissions, and a remote server acts with the access you grant it. Install servers only from their official repo or docs, use read-only modes and narrowly scoped tokens where they exist, keep tool approval on, and remember that a server that reads outside content can carry prompt injection into your session.

Are the original reference MCP servers for GitHub and Postgres still maintained?
No. The MCP project moved its GitHub, GitLab, PostgreSQL, Puppeteer, Redis, Sentry, Slack, SQLite and several other reference servers to an archive of servers that are no longer maintained. Its README points to the current reference servers (Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time) and to the MCP Registry for others.

How do I add an MCP server to Claude Code?
Run claude mcp add --transport http with a name and the server's URL for a remote server, or claude mcp add with a name, then --, then the launch command for a local one. Servers go to local scope by default; add --scope project to write them to a shared .mcp.json file. Type /mcp inside Claude Code to check status and sign in.

What is the difference between a remote and a local MCP server?
A local (stdio) server runs on your computer as a process your client starts, like npx @playwright/mcp@latest. A remote server runs as a web service at a URL, like https://mcp.linear.app/mcp, and usually signs you in with OAuth. Most vendor servers on this list are remote; Filesystem, Git, Playwright and Chrome DevTools run locally.

Ready to build it?
Start with a prompt.

Try building on Greta.sh

left-gradient
left-gradient
Questera Logo
SOC 2 Type II Cert.
SOC 2 Type II Cert.
AI Security Framework
AI Security Framework
Enterprise Encryption
Enterprise Encryption
Security Monitoring
Security Monitoring

© 2026 Questera

Follow in Google