Best MCP Servers in 2026: 12 Picks for Claude Code, Cursor, and Codex

The best MCP servers in 2026 are the ones that turn a general-purpose AI client like Claude Code, Cursor, or Codex into a specialist that can read your docs, edit your repo, drive a browser, query your database, and ship a real app. The Model Context Protocol shipped in late 2024, exploded in 2025, and is now the de facto plug for AI coding agents. Picking the right server for the right job matters more than running ten of them at once.

This guide covers 12 production-ready MCP servers across four categories: documentation and knowledge, developer tooling, operations and infrastructure, and app building. Each pick includes what it does, transports it supports, install effort, and where it fits in a Claude Code, Cursor, or Codex setup. We picked from servers with active maintenance, public install instructions, and either an official vendor backing or significant community adoption (1,000+ GitHub stars or vendor-published).

Quick Answer

• For docs and code references, install Context7 first. It eliminates hallucinated APIs and is the single highest-impact MCP server for any coding agent.
• For repo and CI work, install GitHub MCP (official). PRs, issues, Actions logs, all addressable from your AI conversation.
• For browser automation and visual checks, install Playwright MCP (official from Microsoft). Stealth-mode browsing, screenshots, form fills.
• For shipping a production app from a prompt, install Totalum MCP. It generates and deploys real Next.js apps with auth, payments, database, and storage. Useful for agencies and SaaS teams that want their agent to ship, not just prototype.
• Keep your active MCP set small. Three to six well-chosen servers beats fifteen, because every server adds tool-call latency and tool-name collisions inside the agent's context window.

What is an MCP server, and why this list matters

The Model Context Protocol is an open standard that lets an AI client (Claude Code, Cursor, Codex CLI, Cline, Aider, custom agents) talk to external systems through a uniform tool-call interface. An MCP server is the small program on the other side of that interface. It exposes tools (functions the agent can call), resources (content the agent can read), and prompts (templates the agent can use).

Why does the MCP server you pick matter? Because the agent only knows what its servers expose. A coding agent without Context7 hallucinates package APIs; with Context7 it pulls the real, version-pinned docs. A coding agent without GitHub MCP describes pull requests; with GitHub MCP it actually opens them. A coding agent without an app-building MCP can edit a project; with one it can deploy the project to a real domain.

By 2026 the catalog has hundreds of MCP servers. Most are abandoned demos. The 12 below are the ones we keep installed across our own daily workflows and across the agents we run at Totalum. They are also the ones that show up most often in Claude Code, Cursor, and Codex setups in the public threads we audit each week.

How we picked the 12 servers on this list

Selection criteria, applied in order:

1. Active maintenance in the last 90 days, with releases or merged PRs.
2. Public install instructions that work without a private invite.
3. Real production usage: either an official vendor backing (GitHub, Microsoft, Anthropic, Sentry, Slack, Supabase, Notion, Totalum) or community adoption above 1,000 GitHub stars.
4. Coverage of a unique job-to-be-done, not a duplicate of another server in the list.
5. Transport support for both stdio and HTTP/SSE where relevant. Stdio is faster for local; HTTP is necessary for cloud agents and for IDEs that no longer ship stdio adapters.

Servers ruled out: any server that is a thin wrapper around a single REST endpoint without auth or rate limiting, anything pre-1.0 with no roadmap, anything where the install requires editing source files. Production agents need install steps that fit in a claude_desktop_config.json block.

Documentation and knowledge MCP servers

1. Context7

The single highest-leverage MCP server for any coding agent. Context7 indexes real, version-specific documentation and injects it into the agent's prompt on demand. The agent stops hallucinating package APIs because it can ask for the actual docs of the specific version your project uses.

• What it does: ships up-to-date docs for thousands of libraries, version-pinned, into your prompt.
• Transports: stdio, HTTP.
• Install effort: 5 minutes via npx -y @upstash/context7-mcp.
• Best for: every Claude Code, Cursor, and Codex setup, full stop. If you install only one MCP server, install this one.
• Skip if: you only work in one language with stable APIs and never touch new packages.

2. Firecrawl MCP

Firecrawl converts arbitrary websites into clean markdown that an agent can read. It handles JavaScript rendering, anti-bot pages, and pagination. Great for agents that need to read competitor docs, parse changelogs, or extract structured data from public pages.

• What it does: scrapes any URL into clean markdown, with optional crawl-depth and JSON-extraction modes.
• Transports: stdio, HTTP.
• Install effort: 10 minutes; requires a Firecrawl API key.
• Best for: research-heavy agents, content workflows, SEO audits, and any task where you need to pull live web content into the prompt.
• Skip if: your work is fully internal and the agent never touches the public web.

3. Exa MCP

Exa is a neural search engine optimized for precise web retrieval. The MCP server lets the agent issue semantic queries and get clean results back, which is much better than wrapping a generic search API.

• What it does: semantic web search and content retrieval tuned for AI agents.
• Transports: stdio, HTTP.
• Install effort: 10 minutes; requires an Exa API key.
• Best for: agents that do research, market scans, or competitive intelligence inside a conversation.
• Skip if: you already get research output from a higher-level workflow tool.

Developer tooling MCP servers

4. GitHub MCP (official)

GitHub's official server. The agent can read repos, list branches, open and review pull requests, comment on issues, trigger Actions workflows, and search code across organizations. The official server is more reliable and better-scoped than the community wrappers it replaced.

• What it does: full GitHub API access, scoped by your personal access token or GitHub App.
• Transports: stdio, HTTP/SSE.
• Install effort: 10 minutes; requires a personal access token with the right scopes.
• Best for: any team whose work touches GitHub, which is almost everyone.
• Skip if: you self-host on Forgejo or GitLab; install the equivalent server for those instead.

5. Playwright MCP (official from Microsoft)

Microsoft's browser automation server, built on top of Playwright. The agent can navigate pages, click, fill forms, take screenshots, and run JavaScript inside a real browser. Stealth mode is supported through the underlying Playwright runner. Around 30,000 stars on GitHub by mid-2026.

• What it does: full browser automation through MCP tool calls.
• Transports: stdio.
• Install effort: 5 minutes via npx -y @playwright/mcp.
• Best for: visual testing, UI debugging, scraping JavaScript-heavy pages, agentic web flows.
• Skip if: you only need static HTML; Firecrawl or Exa are lighter for that.

6. Filesystem MCP (official Anthropic)

The reference implementation from Anthropic. Reads, writes, and lists files in directories you allow-list. Boring on paper, indispensable in practice. Pair it with Desktop Commander when you need full terminal control alongside file access.

• What it does: scoped filesystem access.
• Transports: stdio.
• Install effort: 2 minutes; bundled with the standard MCP examples.
• Best for: every local coding agent that edits code outside of an IDE adapter.
• Skip if: your IDE already exposes a sandboxed filesystem to the agent (Cursor and modern Claude Code already do this for the open project, so the standalone server is mostly useful for cross-project work).

Operations and infrastructure MCP servers

7. Kubernetes MCP

Lets the agent inspect and operate clusters. List pods, read logs, describe deployments, apply manifests under tight allow-lists. Indispensable when an agent is the on-call rubber duck for a platform team.

• What it does: kubectl-equivalent operations exposed as MCP tools.
• Transports: stdio, HTTP.
• Install effort: 15 minutes; requires kubeconfig and careful RBAC scoping.
• Best for: SREs and platform engineers using AI agents during incident response.
• Skip if: you do not run Kubernetes, or you forbid agent-driven cluster mutations (in which case the read-only mode is still useful).

8. Sentry MCP

The agent reads errors, traces, and release issues from Sentry. Combine with GitHub MCP and the agent can triage a production error, find the offending commit, and open a fix PR in one conversation.

• What it does: query Sentry projects, errors, and traces.
• Transports: HTTP.
• Install effort: 10 minutes; requires a Sentry auth token.
• Best for: teams already on Sentry that want their AI agent to participate in triage.
• Skip if: you use a different observability stack; pick the equivalent server (Datadog, New Relic, Honeycomb all ship MCP-compatible adapters in 2026).

9. Slack MCP (official)

Lets the agent post messages, read threads, and DM users on Slack. Necessary for any agent whose job is to summarize work, flag incidents, or answer questions in a shared channel. Use it with strict per-channel allow-lists.

• What it does: scoped Slack workspace operations.
• Transports: HTTP.
• Install effort: 20 minutes; requires Slack app install and OAuth scopes.
• Best for: agents acting on behalf of a team in a real workspace.
• Skip if: your team uses Discord, Teams, or Mattermost; pick the matching server.

App building and backend MCP servers

10. Supabase MCP (official)

Supabase shipped an official server in 2025. The agent can query Postgres, run migrations, manage auth and storage, and read functions. Strong fit for teams whose stack is already Supabase.

• What it does: Postgres, auth, storage, and edge function operations on a Supabase project.
• Transports: HTTP.
• Install effort: 10 minutes; requires a project access token.
• Best for: teams already on Supabase who want an agent to ship database changes.
• Skip if: your stack is not Supabase; either use a generic Postgres MCP server or the matching backend's official server.

11. Notion MCP (official)

Notion's official server lets the agent read and write pages, databases, and properties. Excellent for AI agents that maintain a knowledge base, ship docs alongside code, or pull product context into a coding session.

• What it does: page and database operations on a Notion workspace.
• Transports: HTTP.
• Install effort: 10 minutes; requires a Notion integration token.
• Best for: teams whose product context lives in Notion.
• Skip if: your docs live in Confluence or in-repo markdown; use the equivalent there.

12. Totalum MCP

Totalum is the only MCP server in this list that builds and deploys a full production application from inside the conversation. The agent generates a real Next.js project with TotalumSDK, wires up auth, payments, database, file storage, AI integrations, and ships it to a custom domain on Cloudflare. Useful when the goal is a working app, not a code suggestion.

• What it does: create projects, dispatch the Totalum agent to build features, query and edit the project's database, deploy to production domains.
• Transports: HTTP.
• Install effort: 5 minutes; requires a Totalum project and API key.
• Best for: software agencies shipping client work, SaaS teams embedding an AI builder feature, anyone whose AI agent should produce a deployed app rather than only edits to one. See the Claude Code MCP tutorial for a full end-to-end walk-through.
• Skip if: you only want code edits inside an existing repo, with no deploy.

Best MCP servers in 2026 at a glance

#	Server	Category	Transports	Official	Best client	One-line verdict
1	Context7	Docs	stdio, HTTP	community	all	Stops hallucinated APIs. Install first.
2	Firecrawl	Docs	stdio, HTTP	vendor	all	Cleanest web-to-markdown for agents.
3	Exa	Docs	stdio, HTTP	vendor	all	Semantic search tuned for agents.
4	GitHub	Dev tools	stdio, HTTP	official	all	Real repo and CI access.
5	Playwright	Dev tools	stdio	official	Claude Code, Cursor	Browser, screenshots, stealth.
6	Filesystem	Dev tools	stdio	official	local agents	Reference local file access.
7	Kubernetes	Ops	stdio, HTTP	community	platform agents	kubectl through MCP.
8	Sentry	Ops	HTTP	official	triage agents	Errors and traces in-prompt.
9	Slack	Ops	HTTP	official	team agents	Post and read in-channel.
10	Supabase	App	HTTP	official	Supabase teams	Postgres + auth + storage.
11	Notion	App	HTTP	official	Notion teams	Pages and databases.
12	Totalum	App	HTTP	official	builders, agencies	Ship a production Next.js app from a prompt.

How to add an MCP server to Claude Code, Cursor, or Codex

Every modern client follows the same pattern: add an entry to a config file, restart the client, the new tools show up in the agent's tool list. The exact file path differs.

• Claude Code: ~/.config/claude/claude_desktop_config.json (Linux), ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), %APPDATA%/Claude/claude_desktop_config.json (Windows).
• Cursor: Settings, Features, MCP Servers panel. Cursor reads ~/.cursor/mcp.json for global servers and .cursor/mcp.json per project.
• Codex CLI: ~/.codex/config.toml under [mcp_servers]. Codex 1.x ships native MCP support.
• Cline / Aider / Cline-derived agents: each has a settings panel that mirrors Claude Code's JSON shape.

For a step-by-step on Claude Code specifically, follow the Claude Code MCP tutorial. The same patterns transfer to Cursor and Codex.

A minimal config block, applicable to Claude Code, looks like:

{
  "mcpServers": {
    "context7": { "command": "npx", "args": ["-y", "@upstash/context7-mcp"] },
    "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." } },
    "playwright": { "command": "npx", "args": ["-y", "@playwright/mcp"] }
  }
}

Restart the client. Run Tools (Claude Code), MCP (Cursor), or /tools (Codex) to confirm the new tools are loaded.

Common pitfalls and how to avoid them

• Installing too many servers: every server adds tools to the agent's context. Past six or seven, tool selection accuracy drops. Pick the three to six servers that match your daily workflow and disable the rest.
• Tool-name collisions: two servers exposing tools with the same name confuse the agent. Modern clients prefix with the server name, but older versions do not. Check the loaded tool list after install.
• Loose scopes: a GitHub PAT with repo scope, a Slack token with chat:write on every channel, a Kubernetes kubeconfig with cluster-admin. All real risks. Scope down to the minimum.
• Stdio versus HTTP confusion: stdio is faster locally but does not work for cloud-hosted agents. HTTP/SSE works everywhere but adds 30-200 ms latency per call. Pick deliberately, do not pick by default.
• Outdated server pins: npx -y always fetches the latest. For production agents pin a version: npx -y @upstash/context7-mcp@1.4.2. The MCP spec is still evolving in 2026 and breaking changes happen.
• Forgetting the OAuth-only fallback: agents run headless. OAuth-only servers cannot complete the redirect. Use the API-key fallback path that most vendors provide.

FAQ

What are the best MCP servers for Claude Code in 2026?

Context7, GitHub MCP, and Playwright MCP cover 80 percent of Claude Code workflows. For app shipping, add Totalum MCP. For team work, add Slack MCP. The list above ranks all 12 with explicit Claude Code fit.

What are the best MCP servers for Cursor?

Same as Claude Code with one caveat: Cursor already exposes a sandboxed filesystem and project context, so the standalone Filesystem MCP is rarely needed. Start with Context7, GitHub MCP, Playwright MCP, and add an app-building MCP like Totalum or Supabase depending on your stack.

What are the best MCP servers for Codex CLI?

Codex 1.x in 2026 supports MCP natively. The same top picks apply: Context7, GitHub, Playwright. For app shipping from Codex specifically, see the comparison with Claude Code in Claude Code vs Codex.

Are there free MCP servers?

Yes. The official Filesystem and Playwright servers are free. Context7 has a free tier. GitHub MCP is free with a free GitHub PAT. Sentry, Slack, Notion, Supabase, and Totalum MCPs are free with their respective platforms' free tiers; usage beyond those tiers is metered.

Is there a curated list of MCP servers I can browse?

Yes. The official MCP catalog and the community-maintained punkpeye/awesome-mcp-servers list cover hundreds of servers. Use them as a starting point, but do not install more than six at once.

How do I add an MCP server to VS Code?

VS Code has no first-party MCP support yet in 2026. Use the GitHub Copilot-Cline or Continue extensions; both expose MCP server config through a JSON file, similar in shape to Claude Code's. The 12 picks above all work through those extensions.

Should I run my own MCP server?

Only if your workflow has a unique source of context that no public server covers. Most teams are better served running the public servers above and writing thin wrappers around their internal APIs as MCP tools when they actually need them. Premature MCP servers age fast.

Pairing any of these MCP servers with Anthropic's strongest agentic model amplifies what they can do; our Claude Opus 4.7 and Totalum guide shows how Opus 4.7's tool-use reliability lift translates to end-to-end production builds across these tools.

Verdict

The best MCP servers in 2026 are the ones that match the work you do every day. Run Context7 always, GitHub MCP for any repo work, Playwright for browser flows, and one app-building server (Totalum, Supabase, or Notion) if your goal is to ship rather than just edit. Skip the rest until a real workflow asks for them. Add new servers when you find yourself describing instead of doing.

Ready to ship a real app from your favourite MCP client?

Totalum MCP is the only server on this list that takes your AI agent from suggesting code to deploying a real Next.js app on a real domain. Wire it into Claude Code, Cursor, or Codex and your next conversation can produce a production application with auth, payments, and a database. Start free at totalum.app or read the Lovable vs Totalum comparison to see how Totalum differs from prompt-to-prototype tools. For agencies and SaaS teams that want to embed an AI builder in their own product, the Retool vs Totalum comparison walks through the API and MCP surface in detail. If you are an agency evaluating where MCP-driven app building fits in your delivery process, building a client portal without hiring a developer is a worked example of the same workflow.