Best MCP Servers for Cursor: 9 Essential Picks by Use Case

The best MCP servers for Cursor are the few that match the work you do every day: Context7 for current library docs, GitHub for repo work, Playwright or Chrome DevTools for browser testing, Supabase for database tasks, Figma for design context, Linear and Sentry for issues and errors, and Olostep for live web data. In our view, most projects are better served by three to five of them than by twenty.

An MCP server is a small program that exposes tools, such as "query docs" or "open a pull request," to an AI client through the Model Context Protocol. Cursor's agent reads those tool definitions and decides when to call them. For the protocol itself, see Model Context Protocol explained.

AI coding tools are now part of daily work. The 2026 Stack Overflow Developer Survey reports that "the overwhelming majority of respondents now use coding assistants and agents daily (73%)." The open question is which servers make Cursor's agent better, and which only add noise.

How We Chose These Servers

Each server on this list passed four checks. These criteria come before any ranking, so you can apply them to servers we did not cover.

• Official maintainer: the vendor or the MCP project maintains the server. Packages marked archived or "no longer supported," such as the old @modelcontextprotocol/server-github and @modelcontextprotocol/server-postgres npm packages, are excluded.
• Simple transport: the server offers a hosted Streamable HTTP endpoint or a one-line local npx install.
• Scoped access: the server supports OAuth, restricted keys, read-only modes, or project scoping.
• Small or filterable tool surface: the server exposes few tools, or it lets you turn groups of tools off.

This page covers Cursor only. For a client-agnostic list, see our general MCP server roundup.

Why Tool Count Matters More Than Server Count

Tool count matters because every enabled tool gives the agent one more option to read and choose from. Cursor's docs recommend disabling servers you don't need to reduce "tool clutter."

A 2025 RAG-MCP preprint found that filtering tools before the model sees them "significantly cuts prompt tokens (e.g., by over 50%) and more than triples tool selection accuracy (43.13% vs 13.62% baseline) on benchmark tasks." That result comes from the authors' own benchmark, not from Cursor, but it shows the general effect.

You may see claims that Cursor caps you at 40 tools. Cursor's official MCP docs list no fixed tool limit as of October 2026, so treat that number as unverified. The practical rule still holds: switch off servers you are not using. Managing what the model sees is a core part of context engineering.

The Best MCP Servers for Cursor, by Job

The table below compares every pick on the same attributes. All details come from each vendor's official repository or docs, checked on October 7, 2026.

Context7: Current Library Docs

Context7 gives Cursor's agent up-to-date documentation for the libraries in your project. This helps prevent a common failure: the model writing code against an API version that changed after its training data.

Context7 exposes just two tools, resolve-library-id and query-docs. That small surface makes it cheap to leave on globally. The hosted endpoint is https://mcp.context7.com/mcp. It works without an API key at an anonymous rate limit, though Upstash's Context7 repository recommends adding a key.

{
  "mcpServers": {
    "context7": {
      "url": "https://mcp.context7.com/mcp"
    }
  }
}

GitHub MCP Server: Repos, Issues, and PRs

The official github/github-mcp-server lets Cursor's agent read code, manage issues, and review pull requests on GitHub. Its hosted endpoint is https://api.githubcopilot.com/mcp/, and in Cursor the remote server currently uses a personal access token (PAT).

GitHub's server has a large tool surface, so scope it. The X-MCP-Toolsets header enables only the toolsets you list. The X-MCP-Readonly: true header blocks write actions.

{
  "mcpServers": {
    "github": {
      "url": "https://api.githubcopilot.com/mcp/",
      "headers": {
        "Authorization": "Bearer ${env:GITHUB_PAT}",
        "X-MCP-Toolsets": "repos,issues,pull_requests",
        "X-MCP-Readonly": "true"
      }
    }
  }
}

Many older guides still recommend the @modelcontextprotocol/server-github npm package. That package is no longer supported, and its source now sits in the archived MCP servers repository, so use the official server instead.

Playwright MCP: Browser Automation and E2E Checks

Playwright MCP lets the agent open pages, click, fill forms, and check results in a real browser. Microsoft's Playwright MCP repository says the server reads pages through "structured accessibility snapshots, bypassing the need for screenshots or visually-tuned models."

Playwright MCP runs locally through npx @playwright/mcp@latest over stdio. One caveat: Playwright's own README suggests its CLI with skills can be more token-efficient for coding agents than MCP. Use the MCP server when you want the agent to explore a page interactively.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    }
  }
}

Chrome DevTools MCP: Debugging a Live Browser

Chrome DevTools MCP connects the agent to a live Chrome instance for debugging. The Chrome DevTools MCP README says it lets clients "inspect, debug, and modify any data" in that browser, so the agent can check what your code actually does at runtime.

This server exposes about 59 tools across its categories, which is a heavy load for one job. Its --slim mode cuts that to 3 tools. Usage statistics are on by default, and --no-usage-statistics turns them off. Pick Chrome DevTools MCP or Playwright MCP for a given project, not both.

Supabase MCP: Database Work With Guardrails

Supabase MCP connects the agent to your Supabase projects so it can work with the database while you code. The hosted endpoint is https://mcp.supabase.com/mcp, and it signs you in with OAuth.

Two URL options make Supabase MCP safer. read_only=true blocks writes, and project_ref=<id> limits the agent to one project. Supabase's MCP docs warn that "the primary attack vector unique to LLMs is prompt injection," so keep read-only on for anything near production data.

{
  "mcpServers": {
    "supabase": {
      "url": "https://mcp.supabase.com/mcp?project_ref=YOUR_PROJECT_REF&read_only=true"
    }
  }
}

The old Postgres reference server is archived as well and no longer maintained.

Figma MCP Server: Design Context for Frontend Work

Figma's official MCP server gives the agent structured design context from your Figma files. That helps Cursor generate frontend code that matches the design instead of guessing from a screenshot.

Figma offers two versions. The remote server at https://mcp.figma.com/mcp uses OAuth, and Figma's developer docs say it "strongly recommend[s] using the Remote MCP server." The desktop server runs at http://127.0.0.1:3845/mcp and requires the Figma desktop app. Rate limits depend on your Figma plan and seat.

Linear and Sentry: Issues and Production Errors

Linear MCP and Sentry MCP connect Cursor's agent to the work queue and to production errors. Together they let the agent read a bug report, pull the stack trace, and propose a fix in one session.

• Linear MCP: Linear's MCP docs list the endpoint as https://mcp.linear.app/mcp. Linear uses Streamable HTTP as its primary transport and supports OAuth 2.1 or an API key. A /mcp/readonly endpoint blocks writes.
• Sentry MCP: Sentry's MCP server uses the endpoint https://mcp.sentry.dev/mcp, and all connections use OAuth. You can scope the URL to an organization or project, such as /mcp/{org}/{project}.

Olostep MCP Server: Live Web Data Inside Cursor

The Olostep MCP server gives Cursor's agent 10 tools for the live web: scrape a page, search the web, get sourced answers, batch-scrape URLs, crawl a site, and map its URLs. Context7 covers library docs. Olostep covers other public pages, such as a changelog, a pricing page, or a vendor API reference that isn't indexed elsewhere.

The hosted endpoint is https://mcp.olostep.com/mcp with an Authorization: Bearer header. You can also run it locally with npx -y olostep-mcp, which requires Node.js 18 or later.

{
  "mcpServers": {
    "olostep": {
      "url": "https://mcp.olostep.com/mcp",
      "headers": {
        "Authorization": "Bearer ${env:OLOSTEP_API_KEY}"
      }
    }
  }
}

To verify the setup, open Cursor's MCP settings and look for "olostep" with 10 tools. If it shows "Connected, 0 tools," the API key is likely wrong. Full steps are in the Olostep MCP server setup docs. To compare search-focused servers, see our web search MCP comparison.

How to Add an MCP Server to Cursor

You can add an MCP server to Cursor in one click from the Cursor Marketplace, or by editing an mcp.json file. The Marketplace at cursor.com/marketplace lists official plugins with "Add to Cursor" buttons. cursor.directory is a separate, community-run directory.

To add a server manually:

1. Create the config file. Use .cursor/mcp.json in your project root, or ~/.cursor/mcp.json in your home directory for all projects.
2. Add the server entry. Use url (and optional headers) for a remote server, or command, args, and env for a local one.
3. Reference secrets as variables. Write ${env:NAME} instead of pasting keys. Cursor resolves variables in command, args, env, url, and headers.
4. Check the server in Cursor. Open Customize and confirm the server is enabled and lists its tools.
5. Read the logs if it fails. Open the Output panel (Cmd/Ctrl+Shift+U) and select "MCP Logs." If the server still doesn't appear, restart Cursor.

In our experience, common failures are a missing Node.js install, a typo in the JSON, or a bad API key. For every option, see Cursor's MCP documentation, which states that "Cursor supports three transport methods": stdio, SSE, and Streamable HTTP.

Remote vs Local: Choosing a Transport

A transport is how Cursor talks to the server. Pick remote Streamable HTTP by default, and pick local stdio when the data must stay on your machine.

• stdio (local): Cursor starts the server as a process on your computer. You manage credentials yourself. This suits sensitive data and local resources.
• Streamable HTTP (remote): the vendor hosts the server. Cursor connects over HTTP and can sign in with OAuth. Nothing needs to be installed.
• SSE (legacy): Cursor still supports it, but the 2026-07-28 MCP specification states, "The legacy HTTP+SSE transport is also considered to be officially deprecated, with a year-long offramp."

One limit applies: the envFile option only works for stdio servers. Remote servers need their secrets in headers through ${env:NAME} variables.

Project vs Global Configuration

Put servers you use everywhere in the global file, and put repo-specific servers in the project file. Context7, GitHub, and Olostep fit globally. A Supabase project or a Sentry project belongs in that repo's .cursor/mcp.json.

This split also helps control cost. One early Cursor forum request from February 2025 described global servers as "a waste of tokens" and asked for per-project configs, which Cursor now supports. Project files keep each repo's tool list short. You can commit a project file for your team, as long as secrets stay in environment variables.

How to Keep MCP Servers Safe in Cursor

MCP servers can read your code and act for you, so treat each one like a dependency with access to your machine. Three risks matter most:

• Prompt injection: text returned by a tool, such as a web page or a database row, contains instructions the model then follows.
• Tool poisoning: a server's tool descriptions hide malicious instructions.
• Unauthenticated remote servers: a hosted server exposes tools to anyone who finds it.

These risks show up in research. A 2025 study of 1,899 open-source MCP servers found that "7.2% of servers contain general vulnerabilities, and 5.5% exhibit MCP-specific tool poisoning." A May 2026 measurement of remote MCP servers found 7,973 live servers, and "40.55% expose tools without authentication." Both are preprints, and many of the unauthenticated servers were demos, but the pattern supports careful installs.

Practical safeguards:

• Install from the vendor: prefer official servers from the Marketplace or the vendor's own repo.
• Use read-only modes: turn them on for GitHub, Supabase, and Linear unless the agent must write.
• Restrict keys: create API keys and tokens with the fewest permissions the job needs.
• Keep approvals on: Cursor asks for approval before running MCP tools by default.
• Use the allowlist on teams: Cursor's enterprise admins can limit which servers and tools are allowed.

Starter Stacks by Role

A starter stack is three servers chosen for one role. We suggest three servers per role to keep the tool list short.

Start with one stack, use it for a week, and add a server only when the agent hits a task it cannot do.

When You Don't Need an MCP Server

You don't need an MCP server when the task is a fixed script or a one-off fetch. A CLI command or a direct API call runs the same way every time and adds no tool definitions to the agent's context.

MCP fits when the agent must decide whether and when to call a tool during a task. For a deeper look at this trade-off, read MCP vs direct API calls.

FAQ

Is There a Limit on How Many MCP Tools Cursor Can Use?

Cursor's official MCP docs list no fixed tool limit as of October 2026. Still, fewer enabled tools mean less context per request and can improve tool selection.

Where Is the Cursor mcp.json File?

The project file lives at .cursor/mcp.json in your repo root. The global file lives at ~/.cursor/mcp.json in your home directory.

Do I Need to Restart Cursor After Editing mcp.json?

Sometimes. If a new server doesn't appear in Customize, restart Cursor and check MCP Logs in the Output panel.

Does Cursor Support OAuth for MCP Servers?

Yes. Cursor supports OAuth for remote SSE and Streamable HTTP servers, while local stdio servers use credentials you set yourself.

Are Cursor MCP Servers Free?

Every server on this list is free to install. The service behind it, such as Figma, Sentry, or Olostep, may require an account, an API key, or a paid plan.

Will These Servers Work in Claude Code or Windsurf?

Yes, in most cases, because MCP is an open protocol that many clients support. The config file name and location differ between Cursor, Claude Code, and Windsurf.