Setting Up MCP in Cursor (10 Curated Servers)

Cursor supports Model Context Protocol natively. The setup is straightforward — a JSON config file lists the servers, Cursor connects to each on startup, and the tools appear in the chat sidebar and Composer. The harder question is which servers to install and how to scope them.

This page is the setup walkthrough plus the 10 MCP servers we'd install in a Cursor environment. The servers come from our broader best-of-2026 list, filtered for Cursor-specific compatibility and against the realistic constraint that you should run roughly 3 active servers at a time — context-overload is the same problem in Cursor that it is in Claude Code.

The setup

Cursor looks at two mcp.json files:

• ~/.cursor/mcp.json — user-level, applies to every project
• .cursor/mcp.json in a project root — project-scoped

Minimum viable config with one server:

{
  "mcpServers": {
    "github": {
      "command": "uvx",
      "args": ["github-mcp-server"],
      "env": {
        "GITHUB_PAT": "${GITHUB_PAT}"
      }
    }
  }
}

Three things to notice:

For a multi-server config, just add entries:

{
  "mcpServers": {
    "github": { ... },
    "postgres": { ... },
    "playwright": { ... }
  }
}

Cursor reloads the config on next session start. The MCP servers connect lazily — they only start when first invoked, which keeps idle context low.

The 10 servers worth installing

The list, in order of how often we'd actually enable them.

1. github/github-mcp-server (official, post-archive replacement)

The canonical replacement for the archived @modelcontextprotocol/server-github. Read/write GitHub via fine-grained PAT. Detail at /topic/mcp-github.

"github": {
  "command": "uvx",
  "args": ["github-mcp-server"],
  "env": { "GITHUB_PAT": "${GITHUB_PAT}" }
}

Scope the PAT to specific repos. Never commit mcp.json with the PAT inline.

2. crystaldba/postgres-mcp (Postgres, replacement for archived reference)

Postgres MCP server with proper parameterization and read-only mode. Detail at /topic/mcp-postgres.

"postgres": {
  "command": "uvx",
  "args": ["postgres-mcp"],
  "env": {
    "DATABASE_URL": "${DATABASE_URL_READONLY}",
    "READ_ONLY": "true"
  }
}

Read-only by default. Create a dedicated Postgres role for the agent.

3. microsoft/playwright-mcp (browser automation)

Microsoft's official Playwright bridge. Browser navigation, snapshot, click, fill. Tighter sandbox than the archived Puppeteer reference server. Useful for visual review of changes and for testing skills.

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

No auth required (browser runs locally).

4. upstash/context7 (documentation lookup)

Read-only documentation MCP for thousands of libraries (React, Next.js, Prisma, Express, etc.). Helps Claude/Cursor reference current docs instead of training-cutoff information.

"context7": {
  "command": "npx",
  "args": ["-y", "@upstash/context7-mcp"]
}

Read-only by design. Sidesteps the lethal-trifecta concern.

5. brave/brave-search-mcp-server (web search)

Brave Search via MCP. Rate-limited at the API gateway so it can't be cost-blown.

"brave-search": {
  "command": "npx",
  "args": ["-y", "@brave/brave-search-mcp-server"],
  "env": { "BRAVE_API_KEY": "${BRAVE_API_KEY}" }
}

Brave's API has a generous free tier.

6. supabase-community/supabase-mcp (for Supabase-hosted projects)

Supabase REST/RPC wrapper with auth-aware queries. Use only for dev/test environments per Supabase's official caveat — the RLS story through MCP is not yet production-grade.

"supabase": {
  "command": "uvx",
  "args": ["supabase-mcp"],
  "env": {
    "SUPABASE_URL": "${SUPABASE_URL}",
    "SUPABASE_SERVICE_KEY": "${SUPABASE_SERVICE_KEY}"
  }
}

7. makenotion/notion-mcp-server (official Notion)

OAuth-gated, 18 tools, official Notion. Useful when your knowledge base lives in Notion.

"notion": {
  "command": "uvx",
  "args": ["notion-mcp-server"],
  "env": { "NOTION_TOKEN": "${NOTION_TOKEN}" }
}

8. linear.app/docs/mcp (official Linear remote MCP)

Hosted by Linear, OAuth 2.1 with PKCE. Configure the remote endpoint in Cursor:

"linear": {
  "url": "https://mcp.linear.app/mcp",
  "transport": "http"
}

The OAuth flow runs in the browser on first connect.

9. korotovsky/slack-mcp-server (Slack)

OAuth-scoped Slack MCP. Pick this over the official docs.slack.dev server for read-heavy workflows — korotovsky's scopes are more granular.

"slack": {
  "command": "uvx",
  "args": ["slack-mcp-server"],
  "env": { "SLACK_TOKEN": "${SLACK_TOKEN}" }
}

10. cyanheads/obsidian-mcp-server (Obsidian, for note-taking workflows)

Best of three Obsidian implementations. Use Cristian's "mirror vault" pattern from the Obsidian forum — never expose your personal vault directly.

"obsidian": {
  "command": "uvx",
  "args": ["obsidian-mcp-server"],
  "env": { "OBSIDIAN_VAULT_PATH": "${HOME}/notes/mirror" }
}

How to pick 3

You should run 3 at a time, not 10. The 3 you pick depend on the job.

Don't pick by "which MCP server is coolest." Pick by what the agent needs to do.

Cursor-specific considerations

A few details that differ from Claude Code.

Where this fails

What to read next

• /topic/mcp-servers — the broader MCP hub
• /topic/best-mcp-servers-2026 — the curated 18 across all clients
• /topic/mcp-postgres — Postgres MCP deep-dive
• /topic/mcp-github — GitHub MCP deep-dive
• /topic/mcp-security — the architectural risks
• /topic/mcp-tool-overload — why three is the right number
• /topic/cursor-rules-typescript-5-6 — Cursor rules to pair with MCP
• /for/cursor-developers — Cursor-specific resources

Sources

• Cursor. "MCP documentation". The mcp.json schema and Cursor-specific behavior.
• GitHub. github-mcp-server.
• crystaldba. postgres-mcp.
• Microsoft. playwright-mcp.
• Upstash. context7-mcp.
• Brave. brave-search-mcp-server.
• Linear. "MCP documentation".
• Eclipsesource. "MCP Context Overload". Tool-overload measurements.
• Anthropic. "MCP Security Best Practices".
• TrueFoundry. "Best MCP Servers for Claude Code". Companion list (different selections).