In This Article
The Model Context Protocol (MCP) is the most important standard the AI industry has agreed on since the OpenAI Function Calling format. Anthropic published it in late 2024 and by April 2026 it has become the default way AI agents talk to external tools. Cursor, Windsurf, Claude Code, Cline, OpenAI's official tools, and Microsoft's Agent Framework 1.0 all speak MCP. The list of usable servers has grown from a handful at launch to several hundred in 2026.
This piece is the curated list. Not every server in the wild is included. Every server below either has documented stability, broad use, or fills a category gap. Each entry tells you the name, the canonical repository, the install command, the use case, and the maturity tier (Official, Community, Experimental).
What MCP is in one paragraph
MCP is a protocol that lets an AI agent connect to a tool server and call functions on it with structured input and output. The agent sees a list of "tools" the server exposes, calls them with JSON arguments, and receives JSON back. The protocol is transport-agnostic (stdio for local, HTTP+SSE for remote), uses JSON-RPC for the wire format, and includes capability negotiation, session management, and a permissions model. The key win: any MCP-aware agent can use any MCP server without custom integration code. Plug and play.
How to read the maturity tags
Official means published by Anthropic or the canonical product vendor (GitHub for the GitHub server, etc.). Community means a third-party project with broad adoption and active maintenance. Experimental means useful but rough edges, light maintenance, or limited testing.
How to install MCP servers
For Claude Code, the simplest install pattern is to register the server in your .mcp.json at the project or user level. For Cursor, Windsurf, and Cline, the GUI typically guides you through it. The shape of the configuration file is consistent across clients:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/bo/projects"]
},
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxx" }
}
}
}Many servers also support remote (HTTP+SSE) transport. For long-running production setups remote is the better pattern. For local development on a single machine, stdio is faster to set up.
Filesystem and local IO
Filesystem
Read, write, and search files in a sandboxed root directory. The default starting point for almost every project. Path is restricted at start time so the agent cannot escape the configured root.
Repo: github.com/modelcontextprotocol/servers/tree/main/src/filesystem
Install: npx -y @modelcontextprotocol/server-filesystem /your/path
Memory
Persistent knowledge graph for the agent. Lets you save facts, relations, and observations across sessions. Useful for keeping per-project context the agent remembers without re-reading.
Install: npx -y @modelcontextprotocol/server-memory
Sequential Thinking
Tool that gives the agent an explicit reasoning workspace. Improves multi-step planning quality on complex tasks. Lightweight to add to any project.
Install: npx -y @modelcontextprotocol/server-sequential-thinking
Git and GitHub
Git
Local git operations. Status, log, diff, branch, commit. Great for agents that need to inspect repository state without shelling out.
Install: npx -y @modelcontextprotocol/server-git --repository /your/repo
GitHub
Issues, pull requests, repository search, file contents, code search. The most-used MCP server in production. GitHub maintains a remote-MCP server too which avoids token handling on local machines.
Install: npx -y @modelcontextprotocol/server-github or use the official GitHub remote MCP endpoint.
GitLab
Equivalent to the GitHub server for GitLab.com or self-hosted GitLab. Issues, merge requests, project search.
Install: npx -y @modelcontextprotocol/server-gitlab
Communication and project management
Slack
Read channels, post messages, search history. Useful for agents that need to summarize team chatter or post updates on long-running tasks.
Install: npx -y @modelcontextprotocol/server-slack
Notion
Read and write Notion pages, databases, and blocks. Notion ships an official remote MCP endpoint. For knowledge-base workflows this is the right pick.
Install: register the Notion remote MCP URL, authenticate via OAuth.
Linear
Read and create Linear issues, manage cycles, and search projects. Linear publishes an official remote MCP endpoint.
Install: register the Linear remote MCP URL, authenticate via OAuth.
Asana
Tasks, projects, and team management for Asana. Less mature than Linear's official integration.
Install: npx -y mcp-server-asana with API token.
Jira / Atlassian
Jira issue creation, JQL search, comment posting, sprint queries. Atlassian has previewed an official MCP endpoint; community servers cover the gap.
Install: npx -y mcp-server-atlassian
Gmail / Google Workspace
Read inbox, send messages, label management, search threads. Several community servers exist; pick one with active maintenance.
Install: register the Google Workspace MCP server of your choice with OAuth.
Databases and analytics
Postgres
Read-only Postgres access by default. Lets the agent inspect schema, run queries, and report results. Critical for any data-aware coding agent.
Install: npx -y @modelcontextprotocol/server-postgres "postgresql://..."
SQLite
SQLite database server. Read and (optionally) write. Useful for local prototypes and shipped desktop apps.
Install: npx -y @modelcontextprotocol/server-sqlite --db-path /path/to.db
MySQL
MySQL equivalent of the Postgres server. Schema inspection, parameterized queries, read-by-default.
Install: npx -y mcp-server-mysql with connection string.
Snowflake
Snowflake warehouse access for analytical agents. Reads warehouses, runs SQL, returns row sets.
Install: npx -y mcp-server-snowflake with credentials.
BigQuery
Google BigQuery access. Schema introspection, query execution, dataset exploration.
Install: npx -y mcp-server-bigquery with service-account JSON.
Search, browser, and web
Brave Search
Web search backed by Brave. Privacy-respecting, broad coverage, no Google rate-limit issues.
Install: npx -y @modelcontextprotocol/server-brave-search with API key.
Fetch
Fetch a URL, render to Markdown for the agent. The default web-page reader for any agent that browses.
Install: npx -y @modelcontextprotocol/server-fetch
Puppeteer
Headless Chrome automation. Click buttons, fill forms, take screenshots, scrape JavaScript-heavy pages.
Install: npx -y @modelcontextprotocol/server-puppeteer
Playwright
Same idea as Puppeteer with broader browser support. Multiple community implementations; pick the one matching your runtime.
Install: npx -y mcp-server-playwright
Tavily / Exa
AI-native web search APIs. Often higher-quality results for research agents than commodity search APIs. Tavily and Exa each maintain MCP wrappers.
Install: npx -y mcp-server-tavily or mcp-server-exa with API key.
Cloud and infrastructure
AWS
Multiple MCP servers for specific AWS services (S3, Lambda, ECS, CloudWatch). AWS has published several official servers under the awslabs organization.
Install: see github.com/awslabs/mcp
Cloudflare
Cloudflare's official remote MCP server. Workers, KV, R2, D1, DNS, Pages. Useful for managing edge infrastructure from an agent.
Install: register the Cloudflare remote MCP URL, authenticate via OAuth.
Kubernetes
kubectl-equivalent operations from an agent. Useful for SREs who want a structured agent companion. Several community servers; pick one with read-only mode by default.
Install: npx -y mcp-server-kubernetes
Docker
Docker daemon control. List containers, view logs, run commands. Useful for containerized dev environments.
Install: npx -y mcp-server-docker
Developer utilities
Sentry
Pull error events, stack traces, and issue details into the agent. Great for debugging agents that need production error context.
Install: register the Sentry remote MCP URL, authenticate via OAuth.
Stripe
Stripe API access. Products, customers, subscriptions, payments. Test mode by default.
Install: register the Stripe remote MCP URL with restricted API key.
PostHog
Product analytics queries from the agent. Useful for "what did users do this week" style summarization tasks.
Install: register the PostHog remote MCP URL.
Time
Tiny but useful: gives the agent the current time, parses dates, converts timezones. Surprisingly important for any scheduling work.
Install: npx -y @modelcontextprotocol/server-time
Everart
Image generation MCP server with a simple interface. Useful for agents that need to produce illustrations on the fly.
Install: npx -y @modelcontextprotocol/server-everart with API key.
Specialty data and knowledge
Google Drive
Read Google Drive files. Lets agents pull team documents, spreadsheets, and slides into context.
Install: npx -y @modelcontextprotocol/server-gdrive with OAuth.
SAM.gov / federal data
Federal acquisition data: opportunities, contract awards, vendor profiles. A growing space for federal contractors. Maturity varies; verify the source.
Install: see github MCP indexes for "sam.gov" or "federal data".
arXiv
Search and fetch arXiv preprints. Useful for research agents in technical fields.
Install: npx -y mcp-server-arxiv
PubMed
Biomedical literature search and fetch. Useful for healthcare and biotech workflows.
Install: npx -y mcp-server-pubmed
Wikipedia
Search Wikipedia, fetch article text, follow citations. A reliable encyclopedia source for any agent.
Install: npx -y mcp-server-wikipedia
YouTube transcripts
Fetch video transcripts and metadata. Useful for research workflows that include video sources.
Install: npx -y mcp-server-youtube
Security note
MCP servers run with whatever permissions you grant them. A filesystem server pointed at / can read your entire disk. A GitHub server with broad scopes can act as you across every repository. Treat MCP server credentials with the same care you treat any production API key.
Two practical rules. One, scope every credential to the minimum your agent needs. Read-only by default. Production write access only when the workflow requires it. Two, prefer remote MCP servers from the canonical vendor (Notion, Linear, Sentry, Cloudflare) over community-maintained local servers. The vendor's auth flow is hardened in ways a community server is not.
Building your own MCP server
If you have a tool that doesn't have an MCP wrapper, building one is a half-day project. Anthropic publishes SDKs for TypeScript, Python, Java, and Kotlin. The minimum-viable server is about 80 lines of code: declare your tool inventory, accept JSON-RPC calls, return results. The trickier work is testing and packaging, not the protocol itself.
The pattern that works best: build a small server that wraps an internal API, run it locally over stdio first, then promote to a hosted remote MCP endpoint when stable. This is the same arc the official servers followed.
Frequently asked questions
Q: Do I need MCP if I already use Claude Code?
Claude Code ships with built-in tools (Read, Write, Bash, Grep, Glob) that cover local file work. MCP adds external systems: GitHub, Slack, Postgres, Sentry, Notion. Add them as you need them. Most users start with Filesystem and GitHub.
Q: Will MCP work with GPT-5.5 or Gemini 3 Ultra?
Yes. Both OpenAI and Google have shipped MCP support in their tool-using runtimes. Microsoft Agent Framework 1.0 made MCP a first-class citizen. The protocol is now cross-vendor.
Q: Where do I find new MCP servers?
The official server registry at github.com/modelcontextprotocol/servers is the canonical starting point. Search "mcp-server" on npm and PyPI for community packages. Awesome-mcp lists are also useful.
Q: How do I evaluate whether an MCP server is safe?
Three checks. One, who maintains it? Vendor-official is best. Two, does it require excessive permissions? Read-only-by-default is a good signal. Three, does the source code clearly handle authentication and error states? If you cannot read the code, do not install the server.
Q: Can I share MCP servers across projects?
Yes. Configure them at the user level rather than per-project. Use project-level config only for project-specific scopes (a particular Postgres database, a particular GitHub org).
Q: What about MCP servers for federal use?
Federal use of MCP follows the same rules as other AI tools. Verify the underlying service authorization (FedRAMP for SaaS, agency authority for self-hosted), document the data flow, and confirm the agent's data does not leak through MCP into untrusted infrastructure. Public, free MCP servers are not authorized for CUI handling.
Master MCP and Claude Code
Six months. Five U.S. cities. Live instruction on Claude Code, MCP, and production agent workflows. June through October 2026.
See Our Bootcamp