The MCP ecosystem grew from zero in late 2024 to over 10,000 community servers by mid-2026. That's a lot of choice, with very uneven quality. This guide curates the MCP servers actually worth installing — the ones we use ourselves, the ones our clients depend on, and the ones with healthy maintenance and good security practices.
We've organized this by category, with honest opinions on what works and what doesn't. No paid placements — every recommendation here is based on our own usage or careful review.
How we evaluated
For each server, we considered:
- Maintenance: active commits in the last 60 days, responsive issue handling
- Trust: published by Anthropic, the company behind the service, or a known author
- Functionality: tools that work as documented, not stubs
- Security posture: proper authentication, scoped permissions, no telemetry surprises
- Documentation: setup instructions that actually work
The "essential 5" — install these first
If you're new to MCP and want immediate value, start here:
1. Filesystem (Anthropic official)
@modelcontextprotocol/server-filesystem
Read, write, and search files in a directory you specify. Sounds basic, but combined with Claude's reasoning, it becomes powerful. Ask Claude "find all PDFs in my work folder mentioning Acme Corp" and it actually works.
Install:
{
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/work"]
}
}Watch out: scope it to a specific directory. Don't give it full disk access.
2. GitHub (Anthropic official)
@modelcontextprotocol/server-github
Read repos, PRs, issues. Create branches and commits. Comment on PRs. The most useful server for developers — Claude can review your code, draft commit messages, and triage issues.
Install (needs a personal access token):
{
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {
"GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..."
}
}
}Watch out: use a fine-grained token with minimum required permissions. Don't grant org admin if you only need repo read.
3. Slack (community, well-maintained)
@modelcontextprotocol/server-slack
Read channel history, post messages, search across workspaces. Great for "summarize what happened in #product this week" or "find the message about the deploy yesterday."
Watch out: Slack tokens can read a lot. Use bot tokens with limited scopes, not user tokens with admin access.
4. Make.com (official)
Hosted at developers.make.com/mcp-server/. Exposes your Make scenarios as MCP tools.
This is the most accessible path to building custom MCP capabilities without writing code. Build a scenario in Make's visual editor, enable MCP, and it becomes an AI tool. See our Make.com MCP tutorial for setup.
Watch out: operations cost. AI-invoked tool calls consume Make.com operations like any other scenario execution. Budget accordingly.
5. PostgreSQL (Anthropic official)
@modelcontextprotocol/server-postgres
Query a PostgreSQL database. Read-only by default (safer). Useful for "what were our top customers last month?" without bouncing through a BI tool.
Watch out: use read-only credentials. Even with read-only, large queries can lock tables — implement query timeouts on the database side.
Developer tools (advanced)
Cursor / VS Code integration servers
Various community servers expose IDE state to Claude: open files, current cursor position, recent edits. Useful for "explain the function at my cursor" or "find usages of this symbol."
Recommendation: if you use Cursor, just use Cursor's built-in MCP support directly. Don't try to bridge it to Claude Desktop.
Sentry (community)
Read error reports, trends, and stack traces from Sentry. Lets Claude help with incident triage: "what errors spiked in the last hour?" or "summarize the latest fatal errors."
Watch out: respect Sentry's rate limits. High-volume Sentry workspaces can hit limits quickly with AI-driven queries.
Linear (official, by Linear)
Read and create Linear issues, projects, and cycles. Linear publishes their own MCP server, which is well-maintained and feature-complete.
Great for "what's on my Linear roadmap?" or "create a bug ticket about X with high priority."
Productivity tools
Google Drive (Anthropic official)
@modelcontextprotocol/server-google-drive
Read documents, search Drive, list files. Combined with Claude's reading abilities, becomes "ask questions about anything in my Drive."
Watch out: OAuth scopes. Don't grant Drive write access if you only need read.
Notion (Notion official)
Notion has its own MCP server. Read databases, search pages, create entries.
Particularly useful when paired with read access to your team's internal documentation in Notion — Claude can answer "what's our policy on X?" without making up information.
Obsidian (community)
For Obsidian users: read your vault, search across notes, follow backlinks. Several community implementations exist; the one from @obsidian-mcp/server-obsidian is the most actively maintained.
Calendar servers
Google Calendar and Outlook Calendar both have community MCP servers. Quality varies — test before relying on them for scheduling.
The strongest use case: "what's on my calendar tomorrow?" "Find a 30-min slot Bob and I are both free." The AI can answer these questions much better than rule-based calendar bots.
Business tool servers
HubSpot (community, well-maintained)
Read CRM contacts, deals, companies. Some write capabilities (logging notes, updating fields).
Caveat: HubSpot's API rate limits are aggressive. The MCP server passes through these limits, so high-volume usage can hit ceilings.
Salesforce (multiple community implementations)
Salesforce has no official MCP server yet, but several community implementations exist. Quality varies widely. Stick to ones with active maintenance and clear documentation.
Alternative: use Make.com to integrate Salesforce, then expose those scenarios as MCP tools. More reliable than most direct Salesforce MCP servers in 2026.
Stripe (Stripe official)
Stripe ships an official MCP server. Read customers, charges, subscriptions. Limited write capabilities (issue refunds, create invoices) with strong confirmation flows.
This is one of the better official server implementations — clean API surface, good documentation, conservative on dangerous operations.
Intercom (community)
Read customer conversations, search past tickets, look up customer history. Useful for support team augmentation.
Data and analytics servers
SQLite (Anthropic official)
@modelcontextprotocol/server-sqlite
Query any SQLite database. Useful for personal data analysis, prototyping, or exposing read-only access to a SQLite file your application maintains.
Snowflake (community, growing)
Several community MCP servers for Snowflake. Quality is improving fast as enterprises adopt MCP. By end of 2026, expect an official Snowflake MCP server.
Use case: data team augmentation. "What was our revenue by region last quarter?" without needing to log into Snowflake separately.
Memory servers (knowledge bases)
@modelcontextprotocol/server-memory (official) and several community alternatives provide MCP-compliant memory/knowledge base servers. Store facts, retrieve them later, build long-term context.
Underwhelming today (mid-2026), but rapidly improving. Worth watching.
Specialized servers worth knowing
Brave Search (Brave official)
Web search via Brave's API. Privacy-focused alternative to Google for AI-driven web research.
Puppeteer (community)
Headless browser automation as MCP tools. "Take a screenshot of this URL", "navigate this page and extract text". Powerful but slow — not for high-volume use.
Fetch (Anthropic official)
@modelcontextprotocol/server-fetch — simple HTTP fetcher. Hand Claude a URL, get the content back. Useful when you want Claude to read specific pages without granting broader web access.
Agent Firewall (community, security)
Available via Make.com's Library of Agents. Reusable guardrail layer that screens incoming messages for OWASP LLM Top 10 2025 threat categories before they reach your main agents. Worth knowing about for production deployments.
Servers we don't recommend
Being honest: some categories of MCP servers we've seen but don't recommend.
"All-in-one" mega-servers. Servers that try to expose dozens of unrelated APIs in one package. Tend to be poorly maintained, hard to debug, and frequently insecure. Prefer single-purpose servers.
Servers requiring root or system-level access. If a server asks for sudo permissions or unrestricted system access, that's a red flag. Decline.
Servers with no published source code. Closed-source MCP servers are common (vendor-provided), but for community ones, opacity is suspicious.
Older WebSocket-based servers. Some older community servers use non-standard transports. The protocol has standardized on stdio and HTTP+SSE/Streamable HTTP. Avoid the legacy ones — they'll be deprecated.
How to find good MCP servers
Reliable sources for discovering quality servers:
- github.com/modelcontextprotocol/servers — official curated list, includes Anthropic-maintained and community references
- The MCP Discord — active community discussion, recommendations
- The MCP marketplace at glama.ai — third-party directory with ratings
- Vendor documentation — many SaaS companies now list "MCP" in their integrations directory directly
Avoid: random npm packages with no community presence, "AI marketplace" sites that list servers without curation.
What we're watching for late 2026
The MCP ecosystem in 2026 is mid-cycle — past the early-adopter phase, not yet mature. Things to watch:
- Enterprise iPaaS MCP support. Workato, Tray.io, and Mulesoft will likely ship MCP servers in the next 6-12 months. Will reshape the enterprise AI tooling landscape.
- Better authentication. Current MCP auth is workable but feels primitive vs OAuth 2.0 best practices. Expect protocol updates.
- Multi-tenant servers. Most MCP servers are single-user. Patterns for multi-tenant hosting are still emerging.
- Better observability. Tools for monitoring MCP server health, tracking tool invocations, debugging issues. Currently sparse.
Recommendation by use case
If you're a solo developer: Filesystem + GitHub + a custom server you build yourself.
If you're a marketing/operations professional: Make.com MCP + Google Drive + Slack.
If you're a data analyst: PostgreSQL or Snowflake + SQLite + a custom server for your data warehouse.
If you're a sales operations person: HubSpot or Salesforce MCP + Make.com + Slack.
If you're a customer support manager: Intercom or Zendesk + Slack + Notion (for internal docs).
Next steps
- For Claude Desktop config details, see Claude Desktop MCP setup
- For Make.com-specific integration, see MCP with Make.com tutorial
- To build a custom server when nothing fits, see Build an MCP server
- For broader context, return to the MCP complete guide
Frequently asked questions
Are any of these MCP servers free?
All the MCP servers listed are open-source and free to install. You may pay for the underlying services they connect to (GitHub Pro, Make.com, Stripe), but the MCP layer itself has no licensing cost.
Can I install MCP servers without using Claude Desktop?
Yes. Most MCP servers work with any MCP-compatible client: Cursor, VS Code Copilot, ChatGPT (since Q1 2026), or custom-built hosts. Configuration syntax differs by host but the servers themselves are interoperable.
How do I know if an MCP server is safe to install?
Check: published by Anthropic or a verified company, open source with active maintenance, scoped permissions only, no requests for system-level access, clear privacy policy if it sends data to third parties. When in doubt, audit the source code or skip it.
What's the difference between an MCP server and a Claude plugin?
Claude plugins are an older, Anthropic-specific extension model that pre-dated MCP. MCP is the open standard that has replaced plugins. New integrations are built as MCP servers; older plugins are migrating.
Will using many MCP servers slow down Claude Desktop?
Slightly. Each server adds startup time (typically 100-500ms) and uses some memory. With 5-10 servers, the impact is barely noticeable. With 20+ servers, you'll see meaningful slowdown. Disable servers you don't use.
How do I update MCP servers when new versions ship?
For npx-launched servers (with -y flag), npx auto-updates to the latest version on each launch. For globally installed servers, run 'npm update' periodically. For custom servers, follow the publisher's instructions.