Field notes

The MCP servers actually worth installing in 2026

There are somewhere around ten thousand distinct MCP servers in the wild. Registries like Glama and mcp.so list 20,000+ entries once you count wrappers and forks. Most are abandoned, redundant, or exist to check a box on a vendor’s AI story. Meanwhile every server you do install has a standing cost: tool definitions eat context, and each one is another thing that can misbehave.

So this is not a “top 50” list. It’s the short list of servers that, as of mid-2026, earn their keep for a full-stack developer using Claude Code, with the install command and the caveat the vendor’s landing page won’t tell you. (If you’re unsure whether your problem even needs an MCP server, I wrote about when a skill is the better tool.)

One convention up front: nearly every first-party server now uses the same install pattern, claude mcp add --transport http <name> <url>, then /mcp in the session to complete OAuth. When you see that shape below, that’s all there is to it.

1. Context7 — current docs instead of hallucinated APIs

By Upstash · ~59k stars · github.com/upstash/context7

The most popular MCP server in the ecosystem, and for good reason: it injects current, version-specific library documentation into the session, so the model stops confidently writing against an API that changed two majors ago. If you install exactly one server, make it this one.

claude mcp add --scope user --transport http \
  --header "CONTEXT7_API_KEY: YOUR_KEY" \
  context7 https://mcp.context7.com/mcp

The API key is free and optional; it just raises rate limits.

Caveat: the docs are community-contributed with no accuracy guarantee, and only the client is open source (the parsing backend is proprietary). Trust it for API shapes, verify anything load-bearing.

2. Chrome DevTools MCP — let the agent debug what it builds

By Google · ~47k stars · github.com/ChromeDevTools/chrome-devtools-mcp

Full DevTools access from the agent: performance traces, network requests, console output with source-mapped stack traces. This closes the loop that makes frontend work with an agent actually pleasant. It can see the error it just caused.

claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest

Caveat: it exposes everything in the browser to the MCP client and phones home by default. Pass --no-usage-statistics, and --no-performance-crux if you don’t want trace URLs sent to Google’s CrUX API.

3. Playwright MCP — browser automation via accessibility tree

By Microsoft · ~35k stars · github.com/microsoft/playwright-mcp

Drives a real browser through structured accessibility snapshots rather than screenshots: fast, deterministic, no vision model required. Great for exploratory clicking-around and verifying UI changes end to end.

claude mcp add playwright npx @playwright/mcp@latest

Caveat: Microsoft’s own README now steers coding agents toward the Playwright CLI + skills instead, because the MCP tool schemas are token-heavy. Use the server for autonomous browsing; use the plain CLI for routine test runs.

4. GitHub MCP — the official one, carefully trimmed

By GitHub · ~31k stars · github.com/github/github-mcp-server

Issues, PRs, Actions, code scanning, all first-party.

claude mcp add --transport http github https://api.githubcopilot.com/mcp/

Caveat: it ships 70+ tools across 20+ toolsets, which is serious context bloat. Scope it down with --toolsets to what you use. And be honest with yourself: if all you do is create PRs and read issues, the gh CLI does that with zero standing cost.

5. Postgres — pick your flavor

Postgres MCP Pro by Crystal DBA (~3k stars, github.com/crystaldba/postgres-mcp) is the community favorite since Anthropic archived the reference server. It’s not just “run SQL”: EXPLAIN analysis, health checks, and index tuning with hypothetical-index simulation. Run it with --access-mode=restricted for read-only with time limits. Needs Postgres 15+ and the pg_stat_statements / hypopg extensions for the fancy parts.

If your Postgres lives in Supabase, use their official server instead (github.com/supabase-community/supabase-mcp, hosted at https://mcp.supabase.com/mcp): migrations, Edge Function deploys, type generation. Caveat straight from their own README: prompt injection against a production database is a real risk. Use read-only mode and project scoping.

6. Sentry — production errors where the fix gets written

By Sentry · github.com/getsentry/sentry-mcp

Pull real stack traces and root-cause analysis into the session instead of pasting them. The documented install path is now their plugin:

claude plugin marketplace add getsentry/sentry-mcp
claude plugin install sentry-mcp@sentry-mcp

Caveat: the AI-powered search tools need your own LLM provider key configured, and Seer isn’t available self-hosted.

7. Linear — let the agent file its own tickets

By Linear · hosted only · linear.app/docs/mcp

claude mcp add --transport http linear-server https://mcp.linear.app/mcp

Caveat: OAuth cache corruption is a known failure mode. If it stops authenticating, rm -rf ~/.mcp-auth and reconnect.

8. Figma Dev Mode — design-to-code with real data

By Figma · developers.figma.com/docs/figma-mcp-server

Variables, components, and auto-layout data instead of screenshots: the difference between an agent that approximates a design and one that implements it.

claude mcp add --transport http figma https://mcp.figma.com/mcp

Caveat: write-to-canvas is free “during beta”, and Figma has said outright it will become usage-based paid. Don’t build a workflow on the assumption it stays free.

9. Cloudflare — but only the servers you need

By Cloudflare · ~4k stars · github.com/cloudflare/mcp-server-cloudflare

A suite of 12 hosted servers. Three are broadly useful: docs (https://docs.mcp.cloudflare.com/mcp, no auth needed), Workers bindings, and observability/logs.

claude mcp add --transport http cloudflare-docs https://docs.mcp.cloudflare.com/mcp

Caveat: the repo’s own README warns that chained tool calls can blow past context limits. Add the two or three servers you actually use, never all twelve.

Honorable mentions

  • Serena (github.com/oraios/serena, ~26k stars): LSP-backed semantic code navigation and symbol-level editing across 40+ languages. Genuinely powerful on large codebases. The catch is setup: the README explicitly says not to install it via a marketplace, so expect uv tool install serena-agent and some configuration.
  • Firecrawl (github.com/firecrawl/firecrawl-mcp-server, ~7k stars): web scraping into clean, agent-ready markdown. Free keyless tier. Crawls can exceed token limits, so cap depth and page count.
  • Notion (developers.notion.com/docs/get-started-with-mcp, hosted at https://mcp.notion.com/mcp): solid if your specs live there. OAuth-only (no headless use), and even Notion’s docs tell you to keep an eye on /context.

What I deliberately left off

The filesystem reference server tops every generic list (88k stars on the reference monorepo), but for Claude Code specifically it’s redundant: Read, Write, Edit, and Glob are native tools. It matters for Claude Desktop and other clients, not here. The same test applies broadly: does this server do something the agent can’t already do with its own tools or a CLI? If not, skip it.

That test is worth keeping because the costs of over-installing are real and compounding: every server adds standing context weight, another OAuth token to manage, and another third-party tool description your agent implicitly trusts. As the security folks keep demonstrating, that trust is not a theoretical concern. Install few, scope them tight, and check /context once in a while.

Sources

Star counts are approximate, as displayed on GitHub in July 2026.