There are now more than 75,000 MCP servers tracked across the official registry, Glama, Smithery, and PulseMCP. Almost all of them you will never need. This guide is the other side of that number: the 25 servers that actually earn a slot in a working developer setup — what each one does, the real URL, and the exact way to connect it. Every server below works in Cortex (Settings → MCP Servers), Claude Desktop, Cursor, and any other client that speaks the standard mcpServers config.
New to MCP itself? Start with our MCP explainer — this article assumes you know what a server is and want to know which ones to run.
How to connect any of these (30 seconds)
Every server ships one of two ways:
- Local (stdio): the server runs on your machine via a one-line command like
npx -y @upstash/context7-mcp. You paste that command (with any API key it needs) into your client's MCP config. - Remote (hosted): the vendor hosts it — you add a URL like
https://mcp.linear.app/sseand sign in with OAuth. Nothing to install, nothing to update.
In Cortex: Settings → MCP Servers → Add, paste the command or URL, save. The agent picks up the new tools immediately — in chat and inside agentic loops. One practical warning before you connect all 25: every server adds its tool definitions to your context window. Five or six well-chosen servers beat twenty idle ones — connect what the current project needs, disable the rest.
Core coding — the first five to install
1. GitHub MCP
The single highest-value connection for a developer. The agent reads issues, inspects pull requests, searches code across your org, checks CI status, and opens PRs — without you copy-pasting anything into chat. Official, maintained by GitHub, works local (Docker) or hosted. Needs a fine-grained personal access token — scope it to the repos you actually work on.
→ github.com/github/github-mcp-server
2. Context7
Kills the “the model learned this API two versions ago” problem. Context7 injects version-pinned, current documentation for thousands of libraries straight into the prompt — ask about Next.js middleware and it answers from today's docs, not 2024's. Free tier, no key needed to start: npx -y @upstash/context7-mcp.
→ github.com/upstash/context7
3. Playwright MCP
Microsoft's official browser automation server. The agent drives a real browser through the accessibility tree — navigating, clicking, filling forms, reading what rendered — so it can verify the web app it just built instead of claiming it works. Faster and far more reliable than screenshot-based automation. npx -y @playwright/mcp.
→ github.com/microsoft/playwright-mcp
4. Chrome DevTools MCP
Google's official server — and one of the most-starred MCP projects on GitHub. Where Playwright is for driving the page, this one is for debugging it: console errors, network requests, performance traces, DOM inspection, live CSS values. “Why is this page slow?” becomes a question the agent can answer with real trace data.
→ github.com/ChromeDevTools/chrome-devtools-mcp
5. Filesystem MCP
The reference server for scoped file access: you allow-list specific directories, and the agent gets read/write/search inside them and nothing else. Useful for letting a chat client work on folders outside the project — safely. Part of the official modelcontextprotocol/servers repo.
→ github.com/modelcontextprotocol/servers
6. Fetch MCP
The smallest server that pays for itself daily: fetches a URL and converts it to clean markdown for the model. Error pages, changelogs, blog posts, API references — anything public becomes context in one tool call. Also in the official reference repo: uvx mcp-server-fetch.
→ github.com/modelcontextprotocol/servers
7. Memory MCP
A local knowledge graph the agent can write to and query across sessions — decisions, preferences, project facts. Tell it once that your API uses snake_case and it can recall that next week. Reference implementation, stores everything on your machine.
→ github.com/modelcontextprotocol/servers
8. Sequential Thinking MCP
Gives the model a structured scratchpad for multi-step problems: it decomposes, revises, and branches its reasoning as explicit steps instead of one long guess. Noticeably better plans on gnarly refactors. Also from the official reference repo.
→ github.com/modelcontextprotocol/servers
Search & web data
9. Brave Search MCP
Official server from Brave: web, news, image, and video search with a generous free tier (2,000 queries/month) and no tracking. The standard pick for giving an agent general web search. Needs a free API key from Brave.
→ github.com/brave/brave-search-mcp-server
10. Firecrawl MCP
Industrial-strength scraping: JavaScript-rendered pages, pagination, whole-site crawls — all returned as clean markdown. Where Fetch handles one page, Firecrawl handles “crawl these docs and summarize the migration guide.” Needs a Firecrawl API key.
→ github.com/firecrawl/firecrawl-mcp-server
11. Exa MCP
Neural search built for AI consumers rather than human eyeballs: returns dense, relevant passages instead of ten blue links. Strong for research-style questions ("find prior art for this approach"). Official server from Exa Labs, needs an API key.
→ github.com/exa-labs/exa-mcp-server
Databases
One rule for all four: read-only credentials. The agent should inspect schemas and query data, not drop tables.
12. Postgres MCP Pro
The community's chosen successor to the archived reference Postgres server: schema inspection, safe query execution with configurable access levels, index tuning, and EXPLAIN analysis. Writing a migration against a database the agent can actually see beats guessing from your ORM models.
→ github.com/crystaldba/postgres-mcp
13. Supabase MCP
Official server covering the whole Supabase surface: Postgres queries, auth users, storage, edge functions, logs — plus spinning up branches for safe experiments. If your side project runs on Supabase, this turns "check why signups fail" into one instruction.
→ github.com/supabase-community/supabase-mcp
14. MongoDB MCP
Official MongoDB server: query collections, inspect schemas and indexes, and run Atlas admin operations. Read-only mode is a config flag — use it.
→ github.com/mongodb-js/mongodb-mcp-server
15. Redis MCP
Official Redis server — natural-language access to keys, hashes, streams, and pub/sub. Most useful in debugging sessions: "what's actually cached under this prefix right now?"
→ github.com/redis/mcp-redis
Cloud & DevOps
16. Sentry MCP
“Fix the top crash from production” as a one-line instruction: the agent pulls the full issue — stack trace, breadcrumbs, environment, frequency — straight from Sentry and works from real data. Official, hosted (OAuth) or local.
→ github.com/getsentry/sentry-mcp · hosted: mcp.sentry.dev
17. AWS MCP Servers
AWS Labs maintains a whole suite under one repo — documentation lookup, CDK guidance, Cost Explorer, Lambda, DynamoDB, and more. Install only the specific servers you need; the docs server alone (current AWS docs on tap) is worth it.
→ github.com/awslabs/mcp
18. Cloudflare MCP
Official multi-server repo from Cloudflare: Workers builds, KV, R2, D1, DNS, cache analytics, Radar. If your stack deploys to Workers, the agent can read build logs and bindings instead of you describing them.
→ github.com/cloudflare/mcp-server-cloudflare
19. Kubernetes MCP
kubectl-grade cluster access as MCP tools — pods, logs, events, deployments, helm — from the Containers org. "Why is this pod crash-looping?" becomes answerable with live cluster state. Point it at a namespaced, read-mostly kubeconfig.
→ github.com/containers/kubernetes-mcp-server
20. Docker MCP Catalog & Toolkit
Two things in one: manage local containers/images/volumes through the agent, and — more importantly — run other MCP servers inside isolated containers from Docker's curated, signed catalog. The cleanest answer to "how do I run a community server without trusting it with my whole machine."
→ docker.com/products/mcp-catalog-and-toolkit
Team & SaaS
21. Linear MCP
Official hosted server — the agent finds, creates, and updates issues, projects, and comments. Best used for context ("what's in the current cycle?") while you keep triage human. Add https://mcp.linear.app/sse and sign in; nothing to install.
→ linear.app/docs/mcp
22. Notion MCP
Official server for the company wiki: read and write pages, databases, and properties. "What did we decide about auth last month?" stops being a 20-minute search. Hosted (OAuth) or local with an integration token — scope the integration to only the pages it needs.
→ github.com/makenotion/notion-mcp-server
23. Atlassian MCP (Jira & Confluence)
Official remote server connecting Jira, Confluence, JSM, Bitbucket, and Compass via OAuth 2.1 — hosted at mcp.atlassian.com, built with Anthropic as launch partner. The agent reads tickets and writes updates with your exact permissions, nothing more.
→ github.com/atlassian/atlassian-mcp-server
24. Stripe MCP
Official agent toolkit: query customers, subscriptions, invoices, and disputes; search the freshest Stripe docs; create test-mode objects. Connect it in test mode for development — live keys have no business in an agent config.
→ github.com/stripe/agent-toolkit · hosted: mcp.stripe.com
25. Figma Dev Mode MCP
Figma's official server ends the screenshot-the-design workflow: the agent reads real frames — layout, variables, components — and generates code that matches the design system instead of eyeballing pixels. One of the highest-demand MCP integrations of 2026.
→ Figma — Guide to the Dev Mode MCP Server
Bubbling under — worth knowing
- Serena — language-server-powered semantic code navigation and editing for agents.
- Slack MCP — the most-used community Slack server (channels, threads, DMs).
- n8n-MCP — lets the agent build n8n automation workflows from its node documentation.
- Obsidian MCP — your notes vault as agent context.
- Desktop Commander — terminal + file control for chat-only clients (an IDE agent already has this built in).
- Neon MCP — serverless Postgres with database branching via chat.
- Zapier MCP — one server that gateways ~8,000 apps when there's no dedicated server.
The security rules, one more time
- Official or widely-reviewed only. An MCP server is code with access to your data that feeds text into an agent that edits your files. Prefer the vendor's own server; read the code of community ones.
- Least privilege, always. Read-only database users, fine-grained GitHub tokens, test-mode Stripe keys, scoped Notion integrations.
- Treat tool output as untrusted. Cortex's agent loop already does — dangerous commands stay blocked and edits pass the reviewer gate regardless of what a connected tool returns.
Connect the five or six that match your current project, keep credentials scoped, and MCP quietly becomes the most useful thing in your setup. MCP connections in Cortex are part of the Pro subscription; the BYOK IDE itself stays free. Full setup steps are in the docs.
Sources
Frequently asked questions
How many MCP servers exist in 2026?
More than 75,000 are tracked across the official MCP Registry, Glama, Smithery, and PulseMCP — but a working developer setup only needs five or six well-chosen servers.
What are the best MCP servers for coding?
GitHub MCP, Context7 (version-pinned library docs), Microsoft's Playwright MCP, Google's Chrome DevTools MCP, and the official reference servers (filesystem, fetch, memory) form the strongest starter set for day-to-day development.
Do MCP servers slow the AI down?
Every connected server adds its tool definitions to the model's context window, so five or six active servers beat twenty idle ones. Connect what the current project needs and disable the rest.