The Best MCP Servers Worth Installing in 2026

The best MCP servers, without scrolling through 9,800 of them

Open mcpservers.org and the community catalog now lists more than 9,800 servers with no ranking beyond popularity. Most posts that promise a shortcut turn out to be a single vendor promoting its own tool. This list keeps the set small and groups each server by the job it does, with the transport and install detail that decides whether it belongs in your setup.

The best MCP servers for most coding agents are a small, specific set: Context7 for documentation, GitHub MCP for repo work, Playwright MCP or Browserbase for browser automation, Firecrawl for scraping and search, and Supabase MCP for databases. Running three to six of them covers nearly every job without adding tool-call latency or name collisions to the agent's context window.

All of these plug into MCP clients such as Claude Code, Cursor, Cline, and Aider. Each server exposes tools (functions the agent can call), resources (content it can read), and prompts (templates it can use), and the client decides which servers to wire in for a given task.

Docs and knowledge: Context7

Context7 has 61.6k stars, 3k forks, and 959 commits on GitHub. It fetches version-pinned documentation and code examples for a library and puts them directly into the agent's prompt, instead of letting the model guess at an API from training data. In 2026 the project added a Codex plugin with installation instructions (11 June 2026) and shipped production HTTP SDK controls (4 September 2026). The Claude Code plugin now requires authentication on /mcp when the client identifies itself as a plugin, using OAuth or an API key, matching how the Exa MCP plugin gates access; anonymous access on the public URL is unchanged. The team at Totalum, reviewing 12 MCP servers, calls Context7 the single highest-impact install for any coding agent because it stops the agent hallucinating package APIs. It's listed on AI Agents Listing as Context7.

Developer tooling: GitHub MCP

mcpservers.org lists GitHub MCP as an official server for repo search, issues, pull requests, code context, and workflows. Totalum's review puts it plainly: a coding agent without GitHub MCP describes pull requests, and with GitHub MCP it actually opens them. It's a standard first install alongside Context7 for anyone running Claude Code, Cursor, or Codex CLI against a real repository.

Browser automation: Playwright MCP and Browserbase

Playwright MCP is Microsoft's official server, with 36.8k stars, 3.1k forks, and 579 commits. The latest commit lands 3 September 2026, restoring npm publishing through GitHub Actions with OIDC trusted publishing and provenance. Version 0.0.78 added a new browser_find tool that searches the accessibility snapshot of the current page for text or a regular expression and returns matching nodes with surrounding context, cheaper than pulling the full snapshot. The release also distills accessibility snapshots to cut noise and token usage, adds mobile and device emulation through a --mobile flag and a --device option on the CLI, and makes browser_navigate surface non-2xx navigation status codes. Bug fixes in the same cycle enable the chromium sandbox by default, close the isolated HTTP client context on disconnect, and wait for the client event stream before sending server-to-client requests. Both mcpservers.org's community rankings and Totalum's mid-2026 update cite Playwright MCP as the most-used MCP server in the community. It's listed as Playwright Mcp.

Browserbase is mcpservers.org's other official entry for browser automation, described as cloud browser automation rather than a local process. Totalum's review notes that cloud-browser and hosted servers like Browserbase MCP are rising through 2026, the natural pick when you don't want a browser process running on your own machine or CI box. It's listed as Browserbase.

Operations: Firecrawl

Firecrawl MCP is listed on mcpservers.org as an official server for web scraping and search, built for Cursor and Claude. It's the pick when the job is pulling structured content off the web into an agent's workflow rather than driving a browser interactively. It's listed as Firecrawl.

Databases: Supabase MCP

Supabase MCP is an official server that manages Supabase projects, databases, auth, storage, edge functions, and raw SQL. Totalum's review groups it with Slack, Sentry, and Notion as HTTP-first: anything cloud-hosted is moving to HTTP transport rather than stdio. It's listed as Supabase Mcp.

How to choose the best MCP servers for your stack

Start from the job, not the catalog. Totalum's guidance is to keep the active set at three to six servers, because every extra server adds tool-call latency and raises the odds of tool-name collisions inside the agent's context window. Pick one server per job: documentation, repo access, browser automation, and database access cover most coding workflows without overlap. If two servers do the same job, such as Playwright MCP and Browserbase for browser automation, choose by where the browser needs to run, locally or in the cloud, not by installing both.

MCP and Skills are not competing for the same job. Totalum's framing is that MCP handles live tool calls against external systems, while Skills package reusable knowledge the agent loads on demand. A server like Supabase MCP needs a live connection to your database; a skill doesn't need one. AI Agents Listing tracks both kinds of listing separately, alongside the agents that use them, in its MCP Servers directory.

What's new (as of September 2026)

Context7 shipped a Codex plugin with installation instructions on 11 June 2026, added OAuth and API-key gating for the Claude Code plugin, and rolled out production HTTP SDK controls on 4 September 2026. Playwright MCP hit v0.0.78 with the new browser_find tool, distilled accessibility snapshots, mobile and device emulation via --mobile and --device, and non-2xx status surfacing on browser_navigate, with its most recent commit on 3 September 2026. Across the wider catalog, mcpservers.org now counts more than 9,800 servers, and Playwright MCP holds the top spot in its community usage rankings.

Key takeaways

• Totalum's review recommends keeping an active MCP server count between three and six, because each extra server adds tool-call latency and risks tool-name collisions.
• Context7 has 61.6k stars on GitHub and added OAuth-gated authentication for the Claude Code plugin alongside a new Codex plugin in 2026.
• Playwright MCP, with 36.8k stars, is cited as the most-used MCP server in community rankings, and its v0.0.78 release added a browser_find tool for cheaper accessibility-snapshot searches.
• HTTP transport is becoming the norm for cloud-hosted servers, with Totalum grouping Supabase, Slack, Sentry, and Notion together as HTTP-first.
• mcpservers.org's catalog has grown past 9,800 servers with no built-in curation, which is why a short, job-grouped list matters more than the size of the index.

FAQ

What is an MCP server?

An MCP server is an implementation of the Model Context Protocol, an open standard that lets an AI client such as Claude Code, Cursor, Codex CLI, Cline, or Aider talk to external systems through a uniform tool-call interface. It exposes tools the agent can call, resources it can read, and prompts it can use, so the same client can query a database, drive a browser, or search documentation depending on which servers are connected.

Do I need to install every MCP server I can find?

No. Totalum's review is explicit that keeping the active set at three to six servers beats running fifteen, because every additional server adds tool-call latency and increases the chance of tool-name collisions inside the agent's context window. Pick one server per job, documentation, repo access, browser automation, and database access, rather than stacking every server that shows up in a search.

Should I pick Playwright MCP or Browserbase for browser automation?

It depends on where the browser needs to run. Playwright MCP is Microsoft's official server and is community-ranked as the most-used MCP server; it runs as a local process configured with CLI flags like --mobile and --device. Browserbase is described on mcpservers.org as cloud browser automation, and Totalum notes cloud-hosted browser servers are rising through 2026 for teams that don't want the browser process running on their own machine.

What's the difference between MCP servers and agent skills?

MCP servers handle live tool calls against external systems, such as querying a database or driving a browser, while Skills package reusable knowledge an agent loads on demand without a live connection. Totalum's 2026 review frames the two as complementary rather than competing head to head, with Skills picking up jobs that don't need a persistent connection to an outside system.

AI Agents Listing keeps a running set of vetted entries in its mcp-servers hub for readers comparing servers before wiring one into a client.

— The AI Agents Listing team