Best MCP servers for developers: official servers by category

Comparison · MCP

An MCP server is a program that exposes a service's tools and data to AI applications through the Model Context Protocol, and the best MCP servers for developers are the official ones that vendors maintain for their own services. This list covers 15 of them, grouped by the job they do on a software team: code hosting, browsers, documentation, databases, errors, issue tracking, cloud, payments, and web search. A summary table comes first, each entry then gives the same five facts (what it does, who maintains it, transport and auth, best use, and the access to watch), and the last sections cover setup in Claude Code, Cursor, and VS Code, plus security. For background on the protocol itself, start at the MCP topic hub.

Selection criteria#

Every server here is maintained by the company that runs the underlying service, was actively developed in 2026, documents its transport and authentication, and does a job a software engineer needs during development or operations. We checked each server's repository or documentation in November 2026 and left out community wrappers, archived projects, and servers without a stated auth model.

  • Vendor-maintained. The repository or docs page belongs to the service's own organization; popular community servers for the same service are excluded.
  • Active in 2026. Open-source servers had commits in November 2026, and hosted servers have current vendor documentation.
  • Stated transport and auth. Remote servers run over Streamable HTTP (an MCP transport) and name their sign-in method; local servers run over stdio and name the credentials they read.
  • Narrowable access. Servers with read-only modes, scoping parameters, or permission selection at sign-in were preferred.

The MCP project's reference servers, such as Filesystem, Git, and Fetch, are left out because the servers repository calls them educational examples rather than production-ready solutions. Former reference servers for GitHub, Brave Search, and Sentry were archived, and those vendors now maintain their own.

MCP server list: summary table#

The table lists all 15 servers with their maintainer, where they run, how they authenticate, and their category, based on each vendor's repository or documentation as of November 2026. "Remote" means a vendor-hosted Streamable HTTP endpoint, and "local" means a process on your machine, usually over stdio. Several servers offer both.

Server Maintainer Remote or local Auth Category
GitHub MCP Server GitHub Both: api.githubcopilot.com/mcp/ or Docker image OAuth or personal access token Code hosting
Playwright MCP Microsoft Local (stdio; HTTP with --port) None (local browser) Browsers and testing
Chrome DevTools MCP Chrome DevTools (Google) Local (stdio) None (local Chrome) Browsers and testing
Context7 Upstash Both: mcp.context7.com/mcp or @upstash/context7-mcp Optional API key Documentation
Microsoft Learn MCP Server Microsoft Remote: learn.microsoft.com/api/mcp None Documentation
Supabase MCP Supabase Remote: mcp.supabase.com/mcp (subset in the local CLI) OAuth 2.1 Databases and backends
MongoDB MCP Server MongoDB Both: mongodb-mcp-server or Atlas-managed mcp.mongodb.com Connection string; OAuth when managed Databases and backends
Sentry MCP Sentry Both: mcp.sentry.dev/mcp or sentry mcp OAuth; CLI login locally Observability and errors
Linear MCP Linear Remote: mcp.linear.app/mcp OAuth 2.1 or API key Issue tracking
Atlassian Rovo MCP Server Atlassian Remote: mcp.atlassian.com/v2/mcp OAuth 2.1 or API token Issue tracking and docs
Notion MCP Notion Remote: mcp.notion.com/mcp OAuth or MCP token Team docs
AWS MCP Server AWS Remote, through a local signing proxy AWS IAM (SigV4) Cloud
Cloudflare API MCP server Cloudflare Remote: mcp.cloudflare.com/mcp OAuth or API token Cloud
Stripe MCP Stripe Remote: mcp.stripe.com OAuth or agent API key Payments
Exa MCP Exa Both: mcp.exa.ai/mcp or exa-mcp-server None, OAuth, or API key Web search

Local vs remote MCP servers#

A local MCP server runs as a subprocess on your machine, talks to the host over stdio, and reads credentials from environment variables. A remote server runs on the vendor's infrastructure, takes requests over Streamable HTTP, and usually signs you in with OAuth. Most servers on this list are remote; browser and self-managed database servers are the exceptions.

A coding agent host connects to two local servers over stdio on the developer's machine and to two vendor-hosted servers over Streamable HTTP with tokens. LocalRemoteMCP hostClaude Code, Cursor, or VS CodePlaywright MCPstdio subprocesslocal browserMongoDB MCPstdio subprocesskeys in env varsGitHub MCPStreamable HTTPOAuth or PATStripe MCPStreamable HTTPOAuth or agent keyBrowseron your machineDatabaselocal or AtlasGitHub APIStripe APIstdiostdioHTTPSHTTPS
Figure 1: Local MCP servers run as subprocesses with credentials from the environment; remote servers run at the vendor and authenticate with OAuth or a key.

The split decides where credentials live. A remote server's OAuth token is issued for that server and can be revoked from the vendor's settings or your client. A local server inherits whatever its environment holds, so one broad token in an environment variable gives every tool call the same reach. Either way, the model still emits ordinary tool calls, as the MCP explainer describes.

Code hosting: GitHub MCP Server#

GitHub's official MCP server gives an agent access to repositories, issues, pull requests, and Actions, either through a GitHub-hosted remote endpoint or a local Docker container. The MCP project archived its own GitHub reference server, and GitHub's server groups tools into toolsets that you can switch on individually to keep the agent's tool list short.

GitHub MCP Server#

  • What it does: Reads and writes repositories, issues, pull requests, and users. Default toolsets are context, repos, issues, pull_requests, and users; actions, code_security, dependabot, and others are opt-in.
  • Maintainer: GitHub, MIT license; v2.1.0 shipped on November 3, 2026.
  • Transport and auth: Remote at https://api.githubcopilot.com/mcp/ over Streamable HTTP, with OAuth in hosts that register a GitHub App, such as VS Code, or a personal access token (PAT) in an Authorization header elsewhere. Locally it runs over stdio from the ghcr.io/github/github-mcp-server Docker image, with a PAT or a browser login.
  • Best for: Reviewing pull requests, triaging issues, reading code across repositories, and checking CI runs.
  • Caution: A broad token exposes every repository it can reach, so use a fine-grained one. Append /readonly to the remote URL or pass --read-only locally, and pick toolsets with /x/{toolset} or the X-MCP-Toolsets header. The X-MCP-Lockdown header hides public issue content from users without push access, which GitHub calls a best-effort filter, not a security boundary.

Browsers and testing: Playwright MCP and Chrome DevTools MCP#

Two official browser servers cover front-end work: Microsoft's Playwright MCP automates pages through accessibility snapshots, and Google's Chrome DevTools MCP exposes DevTools data such as performance traces, network requests, and console logs. Both run locally, control a real browser, and need no account, which makes the browser profile the main thing to protect.

Playwright MCP#

  • What it does: Automates a browser through Playwright and reads pages through the accessibility tree rather than pixels, so no vision model is needed. Tools include browser_navigate, browser_snapshot, browser_click, and browser_take_screenshot; --caps adds vision, PDF, and DevTools capabilities.
  • Maintainer: Microsoft, Apache 2.0; v0.0.85 shipped on November 9, 2026.
  • Transport and auth: Local: npx @playwright/mcp@latest runs over stdio, and --port switches to HTTP. It has no auth of its own. By default it uses a persistent browser profile that keeps logins between sessions; --isolated keeps the profile in memory.
  • Best for: Reproducing UI bugs, checking a page after a change, and drafting Playwright tests from a live page.
  • Caution: The README says Playwright MCP is not a security boundary, and neither is --allowed-origins. Real logins in a persistent profile become the agent's sessions, so prefer --isolated. For coding agents, the README notes that Playwright's CLI with skills can be more token-efficient.

Chrome DevTools MCP#

  • What it does: Records performance traces and extracts insights, analyzes network requests, reads console messages with source-mapped stack traces, navigates pages, and takes screenshots. --slim exposes a reduced tool set for basic browsing.
  • Maintainer: The ChromeDevTools organization on GitHub, Apache 2.0; v1.13.0 shipped on November 11, 2026.
  • Transport and auth: Local over stdio with npx -y chrome-devtools-mcp@latest, and no auth; --isolated and --headless control the browser instance.
  • Best for: Debugging layout, network, and page-load performance problems with the data a developer would read in DevTools.
  • Caution: The server exposes the browser's content to the MCP client, so keep personal data out of that browser. Usage statistics go to Google by default (--no-usage-statistics turns them off), and performance tools may send trace URLs to the Chrome UX Report API unless you pass --no-performance-crux.

Documentation: Context7 and Microsoft Learn#

Documentation servers give a model current API references at query time, so it is less likely to write code against outdated library versions. Context7 covers many open-source libraries through community-contributed entries, while vendor docs servers such as Microsoft Learn return only the vendor's official documentation. Both are read-only, which makes them the lowest-risk servers on this list.

Context7#

  • What it does: Returns current documentation and code examples for libraries. It has two tools: resolve-library-id finds a library, and query-docs fetches the relevant snippets.
  • Maintainer: Upstash, MIT license.
  • Transport and auth: Remote at https://mcp.context7.com/mcp, or local through the @upstash/context7-mcp package. An API key, sent as a bearer token, is optional and raises rate limits. Upstash also offers a CLI-plus-skills mode (npx ctx7 setup) that skips MCP entirely.
  • Best for: Fast-moving frameworks and SDKs whose APIs changed after the model's training cutoff.
  • Caution: Context7's README says library entries are community-contributed and that it cannot guarantee their accuracy, completeness, or security. Treat returned snippets as untrusted input, like any web content.

Microsoft Learn MCP Server#

  • What it does: Searches and fetches official Microsoft documentation and code samples with three tools: microsoft_docs_search, microsoft_docs_fetch (which returns a page as Markdown), and microsoft_code_sample_search.
  • Maintainer: Microsoft, through the MicrosoftDocs organization on GitHub.
  • Transport and auth: Remote at https://learn.microsoft.com/api/mcp over Streamable HTTP, with no API key, login, or sign-up. The @microsoft/learn-cli package gives terminal access to the same three capabilities.
  • Best for: Azure, .NET, and Microsoft 365 work where a model's knowledge of service names and SDK versions is out of date.
  • Caution: It is read-only and covers only Microsoft content. Its OpenAI-compatible endpoint and maxTokenBudget parameter are marked experimental and subject to change.

Databases and backends: Supabase and MongoDB#

Database servers let an agent inspect schemas, run queries, and apply migrations, so they carry the most risk of any category here. Supabase and MongoDB both maintain official servers with read-only modes and scoping controls. Neon runs a comparable remote server at mcp.neon.tech/mcp with a readonly=true option, though Neon does not recommend using it in production environments.

Supabase MCP#

  • What it does: Manages Supabase projects: list_tables, execute_sql, and apply_migration for the database, deploy_edge_function for Edge Functions, query_logs for debugging, search_docs, and create_branch for development branches. Tools are grouped into feature groups such as database, debugging, development, docs, branching, and storage.
  • Maintainer: Supabase, Apache 2.0; v0.17.0 shipped on November 13, 2026.
  • Transport and auth: Remote at https://mcp.supabase.com/mcp with an OAuth 2.1 login. The Supabase CLI serves a subset of tools at http://localhost:54321/mcp for local development, without OAuth.
  • Best for: Schema changes, migrations, Edge Functions, and log-driven debugging on Supabase projects.
  • Caution: The server runs with your developer permissions, and Supabase warns that data such as a support ticket can carry prompt injection into SQL queries. Scope it with project_ref, add read_only=true for unattended work, limit features, test changes on a branch, and never give the server to end users.

MongoDB MCP Server#

  • What it does: Queries and manages MongoDB data and Atlas: find, aggregate, count, and export for data, atlas-list-clusters and atlas-create-cluster for infrastructure, and atlas-get-performance-advisor for query and index advice.
  • Maintainer: MongoDB, Apache 2.0; v3.3.0 shipped on November 9, 2026.
  • Transport and auth: The local server (npx -y mongodb-mcp-server@latest) runs over stdio or HTTP and reads a connection string from MDB_MCP_CONNECTION_STRING. The Atlas Managed MCP Server, generally available since August 13, 2026, runs at https://mcp.mongodb.com with OAuth through Atlas and works only with Atlas deployments.
  • Best for: Exploring schemas, writing aggregation pipelines against real data, and checking cluster performance.
  • Caution: Start the local server with --readOnly, which allows only read, connect, and metadata operations, or block tools with --disabledTools. Its HTTP mode has no built-in authentication and is meant for localhost. On the managed server, AI client access is off by default for each organization, and the read-only flag is on by default.

Observability and errors: Sentry MCP#

Sentry's MCP server lets an agent read the issues, stack traces, traces, and logs a developer opens when debugging production errors, then propose a fix in the editor. Sentry hosts it as a remote OAuth server and also ships it inside the Sentry CLI for local and self-hosted use, with tools grouped into skills that you can enable or disable.

Sentry MCP#

  • What it does: Groups tools into skills (inspect, seer, triage, and project-management), with natural-language search tools such as search_issues, search_errors, search_traces, and search_logs that translate questions into Sentry queries.
  • Maintainer: Sentry; the code lives in the getsentry/toolkit repository alongside the Sentry CLI.
  • Transport and auth: Remote at https://mcp.sentry.dev/mcp with OAuth. Appending slugs (/mcp/{organizationSlug}/{projectSlug}) scopes the connection to one organization or project. Locally, sentry mcp runs the same server over stdio after sentry auth login, and setting SENTRY_HOST points it at self-hosted Sentry.
  • Best for: Fixing production errors from the stack trace and checking whether a deploy introduced new issues.
  • Caution: Sentry recommends project scoping where possible. Drop skills you don't need with ?disable-skills= on the remote URL or --disable-skills locally, for example project-management.

Issue tracking and team docs: Linear, Atlassian, and Notion#

Linear, Atlassian, and Notion each run a hosted, remote MCP server with OAuth sign-in, so there is nothing to install or update locally. They give agents the context that lives outside the repository: tickets, specs, runbooks, and decisions. Each one acts with the signed-in user's permissions, so the agent can see and change whatever that user can.

Linear MCP#

  • What it does: Finds, creates, and updates Linear issues, projects, and comments.
  • Maintainer: Linear, as a hosted service.
  • Transport and auth: Remote at https://mcp.linear.app/mcp over Streamable HTTP, with OAuth 2.1 and dynamic client registration, or an OAuth token or API key in an Authorization: Bearer header. The /sse endpoint remains only as a deprecated fallback.
  • Best for: Turning an issue into a branch and pull request, and keeping issue status in step with the code.
  • Caution: The default endpoint is read-write. When the agent only needs context, use https://mcp.linear.app/mcp/readonly, request only the read OAuth scope, or create an API key with only the Read permission.

Atlassian Rovo MCP Server#

  • What it does: Connects Jira, Confluence, Jira Service Management, Bitbucket Cloud, Compass, Loom, and other Atlassian apps. The v2 endpoint exposes a discover tool that clients use to find the rest of the tool catalog on demand.
  • Maintainer: Atlassian, as a hosted service.
  • Transport and auth: Remote at https://mcp.atlassian.com/v2/mcp, which Atlassian recommends for new setups, with OAuth 2.1 or an API token for headless use; Jira Service Management takes API tokens only, and Compass takes OAuth only. The legacy /v1/sse endpoint lost support after June 30, 2026.
  • Best for: Teams on Jira and Confluence that want agents to read specs, update tickets, and link work to code.
  • Caution: Access matches what the user can already see in Atlassian Cloud. Admins can restrict which AI tools connect with domain controls and IP allowlisting, and tool usage appears in audit logs. Some tools consume Rovo credits, and Atlassian says most calls cost 1 to 10 credits.

Notion MCP#

  • What it does: Searches a workspace, fetches pages, creates and edits pages, and manages databases and views.
  • Maintainer: Notion, as a hosted service.
  • Transport and auth: Remote at https://mcp.notion.com/mcp over Streamable HTTP with OAuth. On Business and Enterprise plans, a workspace owner can create MCP tokens that clients send as a bearer token, with no browser sign-in. The legacy SSE endpoint retires on February 23, 2027.
  • Best for: Pulling specs, runbooks, and meeting notes into a coding session, and writing results back.
  • Caution: The agent can read and edit anything you can access in the selected workspace. Notion's open-source notion-mcp-server is no longer actively maintained, and Notion's guide calls it a last resort.

Cloud platforms: AWS and Cloudflare#

AWS and Cloudflare have both moved from many product-specific servers to one managed server that covers their whole API. Instead of a tool per endpoint, each offers a few generic tools that can call any API operation, which keeps tool definitions small but means the credential's permissions, not the tool list, define what the agent can do.

AWS MCP Server#

  • What it does: Offers a small, fixed tool set: call_aws runs any of more than 15,000 AWS API operations, run_script runs short Python scripts in a sandbox with no network access, and search_documentation and read_documentation retrieve current AWS docs. AWS service teams maintain Agent Skills for it.
  • Maintainer: AWS; generally available since May 6, 2026, with no additional charge for the server itself.
  • Transport and auth: Remote and managed, with endpoints in N. Virginia (https://aws-mcp.us-east-1.api.aws/mcp) and Frankfurt. MCP Proxy for AWS, an open-source proxy run locally with uvx mcp-proxy-for-aws@latest, signs requests with your IAM credentials, and CloudTrail records the calls.
  • Best for: Inspecting and changing AWS resources without installing a server per service.
  • Caution: call_aws and run_script can do anything the IAM principal can, so use a dedicated role and IAM or service control policies to keep it read-only where possible. The awslabs/mcp repository still publishes local stdio servers per service, and AWS names the Agent Toolkit for AWS as their successor.

Cloudflare API MCP server#

  • What it does: Covers the Cloudflare API with Code Mode: the model writes JavaScript for a search tool that queries the OpenAPI spec and an execute tool that calls the endpoints it found, and the code runs in isolated Workers. A docs tool searches Cloudflare's developer documentation.
  • Maintainer: Cloudflare, Apache 2.0.
  • Transport and auth: Remote at https://mcp.cloudflare.com/mcp, with OAuth (recommended), where you select permissions during authorization, or user and account API tokens for CI. Cloudflare's README puts the Code Mode tools at about 1,100 tokens, against about 244,000 for one tool per endpoint with minimal schemas.
  • Best for: Managing Workers, DNS, and other Cloudflare resources, or debugging a deployment, from one server.
  • Caution: execute can call any endpoint the token's permissions allow, so grant only what the task needs. Older product servers such as Workers Bindings, Workers Builds, and Observability are deprecated in favor of this one; the docs server at docs.mcp.cloudflare.com/mcp remains.

Payments: Stripe MCP#

Stripe's MCP server gives agents read and write access to the Stripe API, plus search over Stripe's documentation and support articles. Stripe hosts it at mcp.stripe.com, supports OAuth with separate permissions for live and sandbox environments, and can require human confirmation in the Dashboard before sensitive actions, such as refunds, go through.

Stripe MCP#

  • What it does: Calls the Stripe API through four generic tools (stripe_api_search, stripe_api_details, stripe_api_read, and stripe_api_write) and searches Stripe's documentation.
  • Maintainer: Stripe, as a hosted service; stripe agent setup in the Stripe CLI installs it along with Stripe's skills.
  • Transport and auth: Remote at https://mcp.stripe.com, with OAuth for interactive clients (separate permissions per live account or sandbox) or an agent API key as a bearer token. Stripe's docs say that from December 2026 the server stops accepting full-access secret keys and restricted keys without the Agent tag, so key-based setups must migrate by November 30, 2026.
  • Best for: Building and debugging payment integrations against a sandbox, and answering support questions about customers, invoices, and subscriptions.
  • Caution: stripe_api_write covers POST, PATCH, PUT, and DELETE. Stripe's Dashboard confirmation applies only to specific actions and is separate from your client's tool approval, so keep both on. Admins can disable MCP access separately for live mode and sandboxes.

Web search: Exa MCP#

Web search servers let an agent look up changelogs, issue threads, and error messages that no documentation server covers. Exa's official server is remote and works without an account, with rate limits. Brave and Tavily also maintain official servers: Brave's runs locally with a Brave Search API key, and Tavily's runs remotely with an API key or OAuth.

Exa MCP#

  • What it does: Searches the web and fetches pages with web_search_exa and web_fetch_exa by default; the ?tools= parameter enables web_search_advanced_exa and agent_run for multi-step research.
  • Maintainer: Exa, MIT license.
  • Transport and auth: Remote at https://mcp.exa.ai/mcp, which works anonymously with rate limits; OAuth or an API key raises them. Exa also publishes the exa-mcp-server npm package for local use.
  • Best for: Research beyond library docs, such as release notes, bug reports, and blog posts about an error.
  • Caution: Every result is third-party web content and a direct path for prompt injection, so keep tool approval on for any write-capable server in the same session.

The web search APIs directory compares the search APIs behind these servers.

How to add MCP servers in Claude Code, Cursor, and VS Code#

Each client stores MCP servers in a JSON config with the same building blocks: a name, then either a URL for a remote server or a command and arguments for a local one. Claude Code adds entries from the terminal with claude mcp add, Cursor reads mcp.json files, and VS Code adds servers through its MCP: Add Server command.

Claude Code MCP setup#

Anthropic's Claude Code MCP documentation recommends the HTTP transport for remote servers and marks SSE as deprecated. For a local server, -- ends Claude Code's own options, so everything after it is the server's command and flags:

# Remote server over Streamable HTTP; run /mcp in a session to sign in
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp
 
# Local stdio server: everything after -- is the server's own command
claude mcp add --transport stdio playwright -- npx -y @playwright/mcp@latest --isolated
 
# Environment variables for a stdio server (keep another option between --env and the name)
claude mcp add --env MDB_MCP_CONNECTION_STRING="$MONGODB_URI" --transport stdio mongodb \
  -- npx -y mongodb-mcp-server@latest --readOnly
 
# Shared with the team through .mcp.json
claude mcp add --scope project --transport http supabase \
  "https://mcp.supabase.com/mcp?project_ref=<project-ref>&read_only=true"

For servers that take an API key instead of OAuth, --header (-H) adds a header such as "Authorization: Bearer <token>". The --scope flag decides where the entry is saved and who gets it:

Scope Stored in Loads in
local (default) ~/.claude.json The current project, for you only
project .mcp.json at the project root The current project, for everyone who clones it; Claude Code asks for approval before first use
user ~/.claude.json All of your projects

A project .mcp.json supports ${VAR} and ${VAR:-default} expansion, so each developer's token stays in their own environment rather than in git:

{
  "mcpServers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/readonly",
      "headers": { "Authorization": "Bearer ${GITHUB_PAT}" }
    },
    "playwright": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@playwright/mcp@latest", "--isolated"]
    }
  }
}

Inside a session, /mcp shows server status and runs OAuth sign-in; claude mcp list, get, and remove manage entries from the shell. Removing a server deletes its stored tokens without revoking them, so run claude mcp logout <name> first. Tool search is on by default, so only tool names and server instructions take up the context window at session start, and MAX_MCP_OUTPUT_TOKENS caps a tool's output at 25,000 tokens by default.

Cursor#

Cursor reads .cursor/mcp.json in a project and ~/.cursor/mcp.json for every project, both with a top-level mcpServers object. Local entries take command, args, and env; remote entries take url and optional headers. According to Cursor's MCP docs, it signs in to remote servers with OAuth using the authorization code flow with PKCE, resolves ${env:NAME} and ${workspaceFolder} variables, and asks for approval before running MCP tools by default.

{
  "mcpServers": {
    "linear": { "url": "https://mcp.linear.app/mcp" },
    "context7": {
      "url": "https://mcp.context7.com/mcp",
      "headers": { "Authorization": "Bearer ${env:CONTEXT7_API_KEY}" }
    },
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest", "--isolated"]
    }
  }
}

VS Code#

In VS Code, run MCP: Add Server from the Command Palette, choose HTTP or a command (stdio), and pick where to save the entry. VS Code's MCP configuration reference recommends the portable mcpServers format for new GitHub Copilot configurations: a workspace-root .mcp.json to share with the team, or ~/.copilot/mcp-config.json for personal servers. The older .vscode/mcp.json file uses a servers key and belongs to VS Code's local, editor-managed mode, which the reference says is deprecated and will be removed in February 2027. Because Claude Code's project scope reads the same root .mcp.json, one file can configure both tools.

Claude Code and VS Code both read a shared .mcp.json in the repository, Cursor reads .cursor/mcp.json, and each client also reads a personal config file in the home directory. Repository, shared through gitHome directory.mcp.jsonmcpServers, committed to git.cursor/mcp.jsonmcpServers, per projectClaude CodeVS CodeGitHub CopilotCursor~/.claude.jsonlocal and user scopes~/.copilot/mcp-config.jsonpersonal servers~/.cursor/mcp.jsonall projectsproject scope
Figure 2: Where each client reads MCP servers from. A root .mcp.json in git serves Claude Code and VS Code; personal entries live in the home directory.

Find more servers in the official MCP Registry#

The official MCP Registry at registry.modelcontextprotocol.io is a metadata catalog for public MCP servers, backed by contributors including Anthropic, GitHub, PulseMCP, and Microsoft. It records where to find each server and how to install it, not the code itself. As of November 2026 it was still in preview, with its v0.1 API frozen since October 2025.

Server names use reverse-DNS namespaces tied to a verified GitHub account or domain, so the namespace is the quickest provenance check. A query of the registry API on November 16, 2026, returned official entries such as io.github.github/github-mcp-server, com.stripe/mcp, com.supabase/mcp, app.linear/linear, and io.github.getsentry/sentry-mcp, alongside servers from other publishers with similar names.

The registry documentation says it is meant mainly for downstream aggregators and delegates security scanning to package registries and those aggregators. The MCP platforms directory compares it with directories such as Glama and PulseMCP.

Security: scope every MCP server you add#

Every MCP server adds tool definitions to the model's context and a new path for actions and untrusted data. Official servers lower the chance of installing a malicious package, but they still return issue comments, web pages, and database rows that can carry instructions. Give each server the least access its job needs.

  • Prefer read-only modes and narrow scopes. Use the controls covered above, from GitHub's /readonly and Supabase's project_ref to permission selection during OAuth for Cloudflare and Stripe.
  • Keep tool approval on for writes. Human-in-the-loop approval in the client is the last check before an agent merges, deploys, refunds, or drops data.
  • Watch for tool poisoning. In tool poisoning, a malicious or compromised server hides instructions in a tool's description or schema. Install from vendor namespaces, pin local package versions in shared configs rather than floating on @latest, and review tool descriptions after updates.
  • Treat tool results as untrusted. Prompt injection can arrive in any tool result, which is why Supabase wraps SQL results in warnings and Stripe cautions against combining its server with others.
  • Keep developer servers away from end users. Supabase says its server runs with developer permissions and should not be given to customers.

When many developers share many servers, an MCP gateway such as Bifrost can put them behind one endpoint and filter which tools each virtual key sees, as the AI gateway explainer describes.

Which MCP servers to install first#

Start with the servers that match the services your team already uses, and add them one at a time. For most web teams, that means GitHub's server, one documentation server such as Context7, one browser server such as Playwright MCP, and the error tracker. Add database, cloud, and payment servers per project, in read-only or sandbox mode first.

For how coding agents use these servers day to day, see the AI coding agents guide and the AI coding tools directory. To expose an internal API the same way, follow how to build an MCP server.

Frequently asked questions

What are the best MCP servers for Claude Code?

For most software teams, start with GitHub's MCP server for repositories and pull requests, Context7 or a vendor documentation server for current library docs, Playwright MCP or Chrome DevTools MCP for browser work, and Sentry's remote server for production errors. Add database, cloud, and payment servers only for the projects that need them, and connect each one in read-only or project-scoped mode first. All of them install with claude mcp add.

How do I add an MCP server to Claude Code?

Run claude mcp add --transport http <name> <url> for a remote server, then type /mcp in a Claude Code session to complete any OAuth sign-in. For a local server, run claude mcp add <name> -- <command> [args], where everything after the -- separator is the server's own command. Add --scope project to save the entry in a .mcp.json file that teammates share, or --scope user to load it in every project.

Where can I find a list of MCP servers?

The official MCP Registry at registry.modelcontextprotocol.io publishes metadata for public MCP servers under names tied to verified GitHub accounts or domains, such as com.stripe/mcp or io.github.getsentry/sentry-mcp. It was still in preview in November 2026. Directories such as the GitHub MCP Registry, Glama, and PulseMCP add search and curation on top, and each vendor's own documentation remains the source for current URLs, auth methods, and flags.

Are MCP servers safe to use?

An MCP server is as safe as its maintainer, its permissions, and the content it returns. Official servers reduce supply-chain risk, but any server can return text that carries prompt injection, such as an issue comment, a database row, or a web page. Prefer read-only or narrowly scoped credentials, keep tool-call approval on for write actions, and connect production data only when a task requires it.

What is the difference between a local and a remote MCP server?

A local MCP server runs on your machine as a subprocess that the host starts and talks to over stdio, and it reads credentials from environment variables. A remote MCP server runs on the vendor's infrastructure and is reached over Streamable HTTP, usually after an OAuth sign-in in the browser. Remote servers need no installation or updates, while local servers can reach your browser, files, and private networks.

How many MCP servers can I connect at once?

The protocol sets no limit, but every connected server adds tool definitions that compete for the model's context window. Claude Code defers MCP tool definitions through tool search by default, loading only tool names and server instructions at session start, and it sets no fixed per-server tool cap. Other clients may load every definition upfront, so enable only the servers and toolsets a project actually uses.

Terms in this guide

Cite this guide

The AI Engineering Editors. "Best MCP servers for developers: official servers by category." The AI Engineering, updated Nov 19, 2026. https://theaiengineering.com/articles/best-mcp-servers

Title
Best MCP servers for developers: official servers by category
Author
The AI Engineering Editors
Publisher
The AI Engineering
Published
Updated
View as Markdown

The AI Engineering Editors ·