mcpPromptFoo Team9 minUpdated Jun 20, 2026

MCP Servers Actually Worth Installing in 2026

A curated list of Model Context Protocol servers that earn their place in a Claude Desktop or Cursor config — with the ones to skip.

What MCP actually gives you

MCP is a standard for connecting LLM apps (Claude Desktop, Cursor, others) to external data sources and tools. Once wired, the model discovers the server's capabilities and can call them like native functions.

The catch: every server you install is another attack surface and another prompt-space consumer. Being selective is the difference between a useful assistant and a bloated one that hallucinates tool calls.

The high-value servers

Filesystem: read/write in a scoped folder. The single most useful MCP server — turns Claude into a real work assistant.

GitHub: PRs, issues, code search. Essential for anyone doing code review or triage in Claude.

Notion or Linear: whichever hosts your work. Read + write access transforms Claude from a chatbot into a workflow assistant.

Postgres (read-only): natural-language analytics against a read-only role. Massive productivity win for PMs and ops.

Situational picks

Slack: powerful for daily-summary use cases; risky if not scoped tightly. Whitelist channels.

Google Drive / Gmail: high value if your work lives there; heavy consent burden and rate limits.

Puppeteer / browser: the killer app for research and scraping; expect flaky runs until server maturity improves.

Time / weather / calculator: cheap, low-risk. Include if you frequently ask the model to compute or reference dates.

The ones to skip

Fetch: overlaps with browsing and often hallucinates URL structure. Prefer a dedicated web search server.

Memory: the model already has session memory; a persistent 'memory' server tends to accumulate noise.

Anything without a well-known publisher. MCP servers run with your local credentials — vet the source before installing.

Security posture

Every MCP server runs on your machine with the tokens you give it. Treat installation like installing a CLI tool with API keys — because that's what it is.

Prefer read-only scopes at first. Elevate to write only for servers whose author you trust and whose behavior you've observed for a week.

Log every tool call for the first month. Claude Desktop's logs are enough; Cursor's are richer. Audit for anything unexpected before moving to production reliance.

FAQs

Related resources