best mcp servers for claude code
The starter lists for this question have a problem: most of what they recommend, Claude Code already does. It reads and writes files, runs shell commands and drives git without any server at all. Adding a filesystem MCP server to it duplicates a native capability and spends context doing it.
So the useful question is narrower: what can it not do, and is that gap worth the context.
Files, shell, and anything reachable from a terminal. That last clause is broader than it sounds - if a capability has a CLI, Claude Code can already use it, and a server wrapping the same CLI adds a layer rather than a capability.
Before installing anything, the honest first question is whether the thing you want has a command line. Very often it does, and you are done.
A browser. The clearest real gap: a terminal cannot render a page, run its JavaScript or click anything. This is the category where a server adds something a shell cannot reach, which is why browser servers are the most installed. Which one depends on the job: choosing an MCP server for browser automation, and if the job is debugging your own site rather than driving one, Playwright MCP vs Chrome DevTools MCP.
A structured view of a live system. A database's schema and query results arrive better through a typed tool than through parsing CLI output, and the same goes for issue trackers with a real API.
Something with no CLI at all. Rarer than people assume, and the actual test of whether a server is worth it.
Tool descriptions travel in the model's context on every turn, not once. Measured on our own server's registry on 2026-09-13, with a tokenizer rather than a characters-per-token rule of thumb: 16 tools, 3,192 tokens resent every turn (9,145 characters of description, plus the argument schemas that travel with them). Two servers of that size on a long coding session is a meaningful slice of the window spent before your question arrives, and it is worse than the tokens: overlapping verbs make the model choose wrongly. How many MCP tools is too many has the arithmetic.
There is also a ceiling you will not be told about. Clients cap how many tools they offer at once, and crossing it silently drops some. If a tool you know exists is being ignored, count your servers before debugging the server.
- One browser server. Not two, ever - the overlap is total.
- Your database, read-only, if the assistant is working against real data.
- Nothing else until something specific fails, because the failure tells you which server you actually needed.
This project's own server is the browser row, and its config block for Claude Code is in the MCP server. It is one entry in one category rather than a recommendation for the set.
What are the best MCP servers for Claude Code? Whichever cover a capability a terminal cannot reach. A browser is the common real one; filesystem servers usually duplicate what it already has.
Do I need a filesystem MCP server with Claude Code? No. It reads and writes files natively.
How many should I install? As few as do the job. Every one costs context on every turn and adds verbs to choose wrongly between.
Are there free MCP servers for Claude Code? Most are open source. The model tokens are billed by your plan either way, and a long agentic session is turn-heavy.
Why is my MCP server not showing up in Claude Code? Usually the config file or the tool ceiling. Check the config block first, then count how many tools all your servers expose together.
See also: how to choose among MCP servers, MCP on GitHub for judging one before installing it, and running AIHawk's browser from Claude Code.
-
The Model Context Protocol documentation, retrieved 2026-09-11, for
tools/listand the shape of a tool declaration. - This project's own MCP server registry, enumerated on 2026-09-13, for the 16 tools and the 3,192 tokens measured above.
Written while maintaining one of the browser servers this page says you should install exactly one of. The first section argues that most starter recommendations are unnecessary, which includes anyone recommending ours for a job a terminal already does.