Best MCP Servers in 2026: 11 Picks for Claude Code and Cursor

The best MCP servers in 2026 are Context7 for current library docs, GitHub's MCP server for repos and pull requests, Playwright MCP and Chrome DevTools MCP for the browser, Serena for code navigation, MCP Toolbox for Databases, Desktop Commander for files and the terminal, Firecrawl and Exa for the web, Framelink MCP for Figma for design, and the Grafana MCP server for observability. Pick the three or four that match the work you do every week, not all eleven.

Every pick below is a dedicated, open-source MCP server with a public repository and an AgenticMart listing. Star counts were read from each GitHub repository on 12 November 2026. Install commands are quoted from each project's own README or docs, and none were run for this post. Each pick also says what it can touch: the files, accounts and outside services it reaches once you connect it.

Top MCP servers compared (stars as of 12 November 2026)

Server	What it does	GitHub stars	Licence	Transport	Listing
Context7	Pulls current, version-specific library docs into your agent	64,684	MIT	Remote (hosted)	/mcp-servers/context7
Chrome DevTools MCP	Drives and inspects a live Chrome: network, console, performance traces	55,160	Apache-2.0	Local	/mcp-servers/chrome-devtools-mcp
Playwright MCP	Browser automation through accessibility snapshots, no screenshots needed	39,558	Apache-2.0	Local	/mcp-servers/playwright-mcp
GitHub MCP Server	Repos, issues, pull requests, Actions and code search	34,413	MIT	Remote or local (Docker)	/mcp-servers/github-mcp-server
Serena	Symbol-level code search, editing and refactoring via language servers	31,327	GPL-3.0-or-later (SolidLSP part MIT)	Local	/mcp-servers/serena
MCP Toolbox for Databases	Prebuilt SQL and schema tools for Postgres, MySQL, BigQuery and more	16,851	Apache-2.0	Local	/mcp-servers/mcp-toolbox
Framelink MCP for Figma	Gives your agent simplified Figma layout and style data	16,197	MIT	Local (calls Figma API)	/mcp-servers/figma-context-mcp
Desktop Commander	Files, search, edits and long-running terminal commands	10,289	MIT	Local	/mcp-servers/desktopcommandermcp
Firecrawl MCP	Search, scrape, crawl and parse web pages into Markdown	7,810	MIT	Local (calls Firecrawl) or remote (hosted)	/mcp-servers/firecrawl-mcp-server
Exa MCP	Web search and page fetching through Exa	5,228	MIT	Remote (hosted)	/mcp-servers/exa-mcp-server
Grafana MCP	Dashboards, datasources, queries and alerts in your Grafana	3,663	Apache-2.0	Local (calls your Grafana)	/mcp-servers/mcp-grafana

"Local" means your client launches the server as a subprocess on your machine, which the MCP spec, revision 2026-07-28 calls the stdio transport. "Remote" means the server runs elsewhere and you connect over Streamable HTTP, where each message is an HTTP POST to one endpoint. A local server can still send your data to an outside API, so read "what it can touch" for each pick. Our guide to local vs remote MCP servers covers the trade-offs.

Stars measure interest, not quality. For the most popular MCP servers by stars across the whole directory, see the MCP servers leaderboard, which you can also download as CSV.

Docs and context

Context7

Good for: stopping your agent from writing code against old or invented APIs. You add "use context7" to a prompt, or name a library ID such as /vercel/next.js, and it fetches current docs and examples for that version. Its two MCP tools are resolve-library-id and query-docs (README).

Install, quoted from the docs, not run for this post. The README offers a one-command setup that needs Node.js 22.13 or newer, with --claude or --cursor to target one client:

npx ctx7 setup

To configure by hand instead, the README gives the server URL https://mcp.context7.com/mcp with your API key in an Authorization: Bearer YOUR_API_KEY header. A free key is recommended for higher rate limits.

What it can touch: it does not read your files. Your library names and questions go to Context7's hosted service. The README says the API backend, parsing engine and crawling engine are private and not part of the repository, and that the indexed docs are community-contributed, so treat fetched snippets as reference material, not instructions.

See Context7 on AgenticMart

Browser

Playwright MCP

Good for: letting an agent click through a web app, fill forms and check pages, using Playwright's accessibility tree instead of screenshots (README). It is a strong default for end-to-end checks and exploratory testing. The README itself notes that coding agents may prefer the separate Playwright CLI with skills, which loads fewer tokens, and that MCP suits longer loops that keep browser state.

Install for Claude Code, quoted from the docs, not run for this post:

claude mcp add playwright npx @playwright/mcp@latest

For Cursor, the README says: go to Cursor Settings, then MCP, then Add new MCP Server, and use the command npx @playwright/mcp@latest.

What it can touch: a browser on your machine and every site it opens, including any session you log into. It has had real bugs: CVE-2025-9611, a DNS rebinding issue, was fixed in version 0.0.40, so keep it updated.

See Playwright MCP on AgenticMart

Chrome DevTools MCP

Good for: debugging rather than clicking. It records performance traces, reads network requests and console messages with source-mapped stack traces, and takes screenshots in a live Chrome (README). Pick it when the question is "why is this page slow or broken", and Playwright when it is "walk through this flow".

Install for Claude Code, quoted from its client configuration guide, not run for this post:

claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest

For Cursor, add the standard config from the README to your MCP settings:

{
  "mcpServers": {
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest"]
    }
  }
}

What it can touch: the README says it exposes the browser's content to your MCP client, which can inspect and modify any data in it. Google collects usage statistics by default (opt out with --no-usage-statistics), and performance tools may send trace URLs to the Chrome UX Report API unless you pass --no-performance-crux.

See Chrome DevTools MCP on AgenticMart

GitHub

GitHub MCP Server

Good for: working with issues, pull requests, branches, Actions runs and code search without leaving your agent. It is GitHub's own server, and the hosted version means nothing to run locally (README).

Install for Claude Code 2.1.1 and newer, quoted from GitHub's Claude install guide, not run for this post:

claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'

For Cursor, quoted from the Cursor install guide, add this to ~/.cursor/mcp.json. The guide says Cursor needs a personal access token for this server and version 0.48.0 or later:

{
  "mcpServers": {
    "github": {
      "url": "https://api.githubcopilot.com/mcp/",
      "headers": {
        "Authorization": "Bearer YOUR_GITHUB_PAT"
      }
    }
  }
}

What it can touch: whatever your token can. A fine-grained token scoped to a few repositories with read-only permissions limits it far more than a classic token with full repo scope. A local Docker version exists if your client cannot reach remote servers.

See GitHub MCP Server on AgenticMart

Code navigation

Serena

Good for: large codebases, where text search and line-number edits get expensive. Serena gives the agent IDE-style tools such as find symbol, find references, rename and replace symbol body, backed by language servers for over 40 languages, or a paid JetBrains plugin (README).

Install, quoted from the README and the client docs, not run for this post. The README asks you not to install it from an MCP or plugin marketplace, because those carry outdated commands:

uv tool install -p 3.13 serena-agent
serena init
claude mcp add --scope user serena -- serena start-mcp-server --context claude-code --project-from-cwd

For Cursor and other IDE clients, the docs recommend launching serena start-mcp-server with --context ide to cut down on tools that duplicate the editor's own.

What it can touch: your project's source, which it reads and edits, plus a memory store it keeps for long-running work. Its basic tools include execute_shell_command; the README says these overlap tools are usually off by default inside harnesses such as Claude Code. Note the licence: the Serena application is GPL-3.0-or-later.

See Serena on AgenticMart

Database

MCP Toolbox for Databases

Good for: querying and exploring databases in plain language. With a --prebuilt=<database> flag you get tools such as list_tables and execute_sql for PostgreSQL, MySQL, SQL Server, BigQuery, Spanner, MongoDB, Redis, Snowflake and others, and a tools.yaml file lets you define narrower custom tools (README).

Install, quoted from the README, not run for this post. Add this to your client's MCP config (for Cursor, mcp.json), then set the connection environment variables from the prebuilt tools reference:

{
  "mcpServers": {
    "toolbox-postgres": {
      "command": "npx",
      "args": [
        "-y",
        "@toolbox-sdk/server",
        "--prebuilt=postgres",
        "--stdio"
      ]
    }
  }
}

For Claude Code, the README lists it as a supported client but gives no CLI one-liner; how to install an MCP server shows how to add a stdio server like this one.

What it can touch: every table the database user can reach. execute_sql runs what the model writes, so connect with a read-only role, or use a narrower toolset such as --prebuilt=postgres/data, or define custom tools in tools.yaml.

See MCP Toolbox on AgenticMart

Filesystem and terminal

Desktop Commander

Good for: an agent that needs to read and edit files, search folders, work with Excel, PDF and Word files, and run long terminal processes with paged output (README). It is built on the reference MCP filesystem server and adds search-and-replace editing and process control. It is most useful in clients without their own file and shell tools.

Install for Claude Code, quoted from the README, not run for this post:

claude mcp add --scope user desktop-commander -- npx -y @wonderwhy-er/desktop-commander@latest

For Cursor, add the README's standard config to ~/.cursor/mcp.json or .cursor/mcp.json in your project:

{
  "mcpServers": {
    "desktop-commander": {
      "command": "npx",
      "args": ["-y", "@wonderwhy-er/desktop-commander@latest"]
    }
  }
}

What it can touch: your file system and your shell. The README says its guardrails are "not a sandbox", and that allowedDirectories limits file tools but terminal commands can still reach files outside those folders. It collects pseudonymous telemetry, which the README says excludes file contents, paths and command arguments; telemetryEnabled turns it off.

See Desktop Commander on AgenticMart

Web search and scraping

Firecrawl MCP

Good for: turning web pages into clean Markdown or structured data. It can search, scrape a URL, crawl or map a site, parse PDFs and interact with dynamic pages (README).

Install, quoted from the README, not run for this post. Run locally with an API key:

env FIRECRAWL_API_KEY=fc-YOUR_API_KEY npx -y firecrawl-mcp

For Cursor, the README's config is:

{
  "mcpServers": {
    "firecrawl-mcp": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": {
        "FIRECRAWL_API_KEY": "YOUR-API-KEY"
      }
    }
  }
}

The README also lists a hosted keyless endpoint, https://mcp.firecrawl.dev/v2/mcp, which exposes three tools.

What it can touch: the URLs you ask it to fetch, sent through Firecrawl's service unless you self-host. Scraped pages are untrusted text that can carry instructions aimed at your agent, so keep approval on for tools that act on what it reads.

See Firecrawl MCP on AgenticMart

Exa MCP

Good for: fast web search and page fetching with little setup. The default tools are web_search_exa and web_fetch_exa, and you can enable advanced search and multi-step research agents through the URL (README).

Install, quoted from the README, not run for this post. For Claude Code:

claude plugin install exa@claude-plugins-official

For Cursor and other clients, point them at the hosted server (the README also has a one-click Cursor install link):

{
  "mcpServers": {
    "exa": {
      "type": "streamable-http",
      "url": "https://mcp.exa.ai/mcp"
    }
  }
}

What it can touch: no local files. Your search queries and fetched URLs go to Exa, and the research agent tool needs OAuth or an API key.

See Exa MCP on AgenticMart

Design

Framelink MCP for Figma

Good for: building UI from a Figma link. You paste a link to a file, frame or group, and the server fetches it from the Figma API, then trims the response to the layout and styling details the model needs (README). The README says it is designed for Cursor, which makes it one of the best MCP servers for Cursor users who work from designs, though any MCP client can run it.

Install on macOS or Linux, quoted from the README, not run for this post (Windows wraps the command in cmd /c):

{
  "mcpServers": {
    "Framelink MCP for Figma": {
      "command": "npx",
      "args": ["-y", "figma-developer-mcp", "--figma-api-key=YOUR-KEY", "--stdio"]
    }
  }
}

What it can touch: every Figma file your personal access token can read. The README also allows setting FIGMA_API_KEY in the env field, which keeps the key out of the arguments list.

See Framelink MCP for Figma on AgenticMart

Observability

Grafana MCP

Good for: asking about incidents in plain language: search dashboards, read panel queries, query datasources, and work with alerts (README).

Install, quoted from the README, not run for this post. It needs uv, and the README names both Claude Desktop and Cursor for this config:

{
  "mcpServers": {
    "grafana": {
      "command": "uvx",
      "args": ["mcp-grafana"],
      "env": {
        "GRAFANA_URL": "http://localhost:3000",
        "GRAFANA_SERVICE_ACCOUNT_TOKEN": "<your service account token>"
      }
    }
  }
}

What it can touch: your Grafana instance, as far as the service account allows, including creating and updating dashboards. The README documents a --disable-write flag for read-only use, and warns that raw SQL query tools run whatever query they get, which is why read-only mode removes them.

See Grafana MCP on AgenticMart

Best MCP servers for Claude Code and for Cursor

The best MCP servers for Claude Code are the ones that add something it does not already have. Claude Code reads files and runs shell commands on its own, so Desktop Commander adds little there. Context7, GitHub, Playwright or Chrome DevTools, and Serena fill real gaps. Every one of those has a claude mcp add (or add-json) command quoted above.

The best MCP servers for Cursor follow the same logic. Cursor also edits files, so start with Context7 and GitHub, add Framelink for Figma if you build from designs, and add Toolbox if you query databases. Every pick above works with a Cursor mcp.json entry, and several (Context7, GitHub, Chrome DevTools, Desktop Commander, Exa) offer one-click Cursor install links in their READMEs.

A starter set for most developers: Context7, GitHub MCP Server and one browser server. Add the rest when a task calls for them. Each server adds tool descriptions the model has to read, and each one is another thing with access to your accounts.

How we picked

We started from the AgenticMart directory, using the live search API and the MCP servers leaderboard, and kept a server only if it met all of these:

1. It is a dedicated MCP server. We left out awesome lists, SDKs and frameworks, and apps that merely support MCP among other features. That is why some high-star leaderboard entries are not here.
2. It covers a common job. One or two picks per job: docs, browser, GitHub, code, database, files, search, design and observability.
3. It has an active public repository with an open-source licence. We read each licence from the repository on 12 November 2026.
4. It has clear install docs for Claude Code, Cursor or a standard mcpServers config, so you can install from its own docs.
5. Where we had a choice, we preferred the vendor's own server (GitHub, Microsoft, Google, Grafana, Firecrawl, Exa) over a third-party wrapper.
6. Stars break ties, not the other way round. Exa and Grafana have fewer stars than many leaderboard entries, but they were the highest-starred vendor servers for search and observability that we found in our directory.

What we did not do: we did not run, benchmark or audit these servers for this post. AgenticMart listing pages show what each one can touch, its install commands and, where one exists, a dated scan report. A scan describes what it looked for on that date. It is not a guarantee.

FAQ

What are the best MCP servers in 2026?

By job: Context7 for docs, GitHub MCP Server for repositories, Playwright MCP or Chrome DevTools MCP for the browser, Serena for code navigation, MCP Toolbox for databases, Desktop Commander for files and the terminal, Firecrawl or Exa for the web, Framelink for Figma, and Grafana MCP for observability. The table above has dated stars, licences and links.

What are the best MCP servers for Claude Code?

Context7, GitHub MCP Server, Playwright MCP or Chrome DevTools MCP, and Serena. Claude Code already has file and shell tools, so servers that bring outside context or IDE-level code navigation add the most. Each has a quoted claude mcp add command above.

What are the best MCP servers for Cursor?

Context7 and GitHub MCP Server, plus Framelink MCP for Figma if you implement designs and MCP Toolbox if you work with databases. All of them install through a Cursor mcp.json entry, and several have one-click install links in their READMEs.

What are the most popular MCP servers?

By GitHub stars on 12 November 2026, Context7 (64,684), Chrome DevTools MCP (55,160), Playwright MCP (39,558) and GitHub MCP Server (34,413) lead this list. The leaderboard ranks every listed server by stars, and you can download it as CSV.

Is it safe to install MCP servers?

That depends on what a server can touch and what you connect it to. A server runs with whatever access you give it, and text it reads can try to steer your agent. Use narrow tokens and read-only modes where offered, keep tool approval on, and read our MCP server security guide before connecting anything to production.

Where can I find a full MCP servers list?

Browse MCP servers on AgenticMart to filter by category, or start with what an MCP server is if you are new. If you run many servers across a team, MCP gateways can centralise access and logging.

Built an MCP server?

If your server does one job well, put it where developers choose what to install. On AgenticMart you can list free now, paid selling opens soon. List your MCP server and it gets a listing page with install commands and a "what it can touch" summary.