Originally published in Spanish on El Rack, a Spanish tech review site I write for. Browser translation handles the rest of the site well if you want to dig deeper into the sysadmin/homelab content.
Claude Code is my main tool every day for managing a handful of self-hosted projects. So when "just install these 10 MCP servers" guides kept showing up in every forum thread, I wanted to actually separate what's worth the setup time from what's just noise sitting in your context window.
Here's what I found after digging into the current state of the ecosystem.
The cost nobody mentions before you install ten of these
Before the list: the one number that matters most. Every connected MCP server injects somewhere between 2,000 and 5,000 tokens of tool schemas at session start. Three servers connected at once is already 6,000 to 15,000 tokens spent before you type your first prompt.
That's not free, even though it feels like it. If Claude Code feels slower to start or seems to "forget" context earlier than you'd expect, check how many MCP servers you're leaving connected by default.
Correction (added after publishing): Claude Code now enables Tool Search by default, which defers MCP schema loading until a tool is actually used — a connected-but-idle server now costs closer to a name-and-description stub than a full schema dump. The 2,000–5,000 token figure above was accurate before this became the default, and still applies if you disable Tool Search or if a given server doesn't support deferred loading yet. Thanks to a reader for catching this in the comments.
MCP is convenient. It is not free. Every server you connect is tokens that are no longer available for your own code.
The three actually worth installing first
Independent guides keep converging on the same top three, and after using them daily, I agree with the consensus:
Context7 — live library documentation, so Claude doesn't hallucinate APIs that changed three versions ago. This is the one that's saved me the most from confidently-wrong code suggestions based on outdated library docs.
Official GitHub MCP — repo operations, pull requests, issues and CI, without leaving the chat. One caveat on the token cost above: for simple lookups, the gh CLI is still much cheaper than going through MCP.
Playwright MCP — lets Claude drive a real browser to confirm a frontend change actually renders correctly, instead of just trusting the code diff. If you do any frontend work, this is the difference between "the code compiles" and "the code actually works."
Two that Anthropic still maintains but aren't essential anymore
The original reference servers for structured reasoning and file operations are still around, but the general consensus is honest about it: native model reasoning improved enough through 2026 that most people skip the reasoning server entirely, and Claude Code already ships built-in file tools that make the equivalent MCP mostly redundant.
The governance shift that matters long-term
In December 2025, MCP governance moved from Anthropic alone to the Agentic AI Foundation under the Linux Foundation, with Anthropic, OpenAI, Google, Microsoft, AWS, Cloudflare and Bloomberg on the board. In practice, this means vendor-maintained servers (Slack's own MCP, GitHub's own MCP) tend to be more reliable than generic third-party integrations — each company simply knows its own service better than any middleman.
Best MCP for a specific task
If you don't want to read the full ranking, here's the direct answer for whatever you're trying to solve:
Working with your database → The PostgreSQL MCP if you run Postgres on any provider, or the official Supabase MCP if your backend lives there specifically (you can scope it to a single project via a URL parameter — good security practice).
Repo management → The official GitHub MCP, already covered above.
Verifying frontend actually works → Playwright, also covered above.
Project docs and knowledge base → The Notion MCP, if your team documents there — keeps pages synced with code changes automatically.
Task management → The Linear MCP, the most widely adopted for this.
Production error monitoring → The Sentry MCP — brings real production error context straight into the chat instead of copy-pasting logs by hand.
Design → The Figma MCP — pulls measurements, components and assets directly from published designs.
Notifying a team → The Slack MCP — post deployment updates or task results without leaving the session.
The question isn't "which MCP servers exist" — it's "which task do I repeat constantly that could be solved without switching windows." That question gets you a better answer than any generic ranking.
When to use MCP and when not to
The rule that sums it up best: use a CLI when latency and tokens matter and you can make a direct shell call; use an MCP server when you want the same tool available across Claude Code, Cursor and other clients with one config; use a Skill when the workflow is a set of instructions, not an external tool to call.
Full article (in Spanish) with more detail: elrack.es/herramientas-ia/mejores-mcp-claude-code-2026
Top comments (6)
The 2,000-5,000 tokens per connected server figure matches what I have measured, but one recent change shifts the calculus: recent versions of Claude Code defer MCP tool schemas and load them on demand via tool search, so a connected-but-unused server costs closer to a name in a list than a full schema dump. The trade moves from startup context tax to one extra round-trip on first use. That weakens the strongest argument for disconnecting the GitHub MCP in favor of gh CLI, though for simple lookups the CLI is still cheaper per call. On Context7: have you compared it against just letting the agent fetch the library's own docs page? For fast-moving libraries I have found live-fetch of the changelog beats a docs index that is itself a few days stale, curious if your experience differs for the self-hosted stack.
Good catch, and you're right — I should have been clearer here. Tool Search is on by default in current Claude Code versions, so a connected-but-idle server now costs closer to a name-and-description stub than a full schema dump. That does weaken the "disconnect what you're not using" framing I had for the token-cost section.
Where I'd still push back slightly: for a single simple lookup (e.g., "what's the status of PR #42"), you're now trading upfront context cost for a round-trip on first use plus whatever the search itself costs in tokens — so gh pr view 42 is still cheaper per call in that narrow case. But you're right that "every connected server taxes every turn" is no longer accurate as a blanket statement, and I'll update the piece to reflect that.
On Context7 vs. live-fetching a library's own docs/changelog: I haven't run a rigorous side-by-side, so take this as anecdotal rather than a real comparison — for anything moving fast (weekly releases, changelog-driven), I'd trust your instinct that a live fetch beats an index that's even a few days stale. Where Context7 has felt more useful for me is less-actively-maintained libraries where the docs site itself is inconsistent or poorly structured for an LLM to parse cleanly, and a normalized index does better than scraping raw HTML. For the self-hosted stack specifically I haven't tested this rigorously enough to give you a confident answer — will actually run that comparison before I claim one way or the other next time I touch it.
If you're doing anything web-related, give MCP-Lite a try, your wallet will thank you later
Thanks for the pointer — I found a couple of different projects called "mcp-lite" (one's a server-building framework, another's a generic client) but nothing that obviously matches "web-related, saves money" from what I could find. Got a link to the specific one you mean? Happy to actually try it before recommending it to anyone else.
GitHub
And my article. Essentially, MCP-Lite ditches DOM parsing for the AOM, which cuts the amount of tokens you need to process by ALOT.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.