Home › How many MCP servers is too many?

How many MCP servers is too many?

Updated 2026-10-23

There is no fixed number. The limit is context: most clients load every tool definition before your first message. Anthropic measured five common servers at about 55K tokens for 58 tools. Load only what today's task needs, prefer servers with toolsets, and turn on tool search where your client offers it.

Want the short version? Get the one-page MCP Context Cheat Sheet.

This page has no affiliate links. How we make money: our disclosure.

How many MCP servers do I need?

Fewer than most setups carry. Start from the task, not the catalog: one server per system the agent must actually read or change today, such as your code host, your issue tracker and one data source. Anthropic's own guidance says tool search pays off once tool definitions pass about 10K tokens or you run 10 or more tools, which a single large server can reach alone. A business run by agents does better with a small fixed set per job than one agent holding every server. Fewer tools also stretch Grok Bot usage limits. Our Crew Builder template splits work into small bots, each with its own short tool list. Start from our best MCP servers list. See real agent setups and the servers each one runs.

We may be paid by X based on how many people use our Crew Builder template.

Why does every MCP server cost context?

New to this? See what an MCP server is. Most MCP clients load all tool definitions upfront: each tool's name, description and input schema goes into the model's context before your request. Anthropic's five-server example (GitHub, Slack, Sentry, Grafana, Splunk) adds up to 58 tools and about 55K tokens, and it reports seeing tool definitions reach 134K tokens internally before optimization. Tool results cost context too: a two-hour meeting transcript passed between tools could add about 50,000 tokens.

Which servers are heaviest?

In Anthropic's count, GitHub's server used about 26K tokens for 35 tools and Slack's about 21K for 11 tools, while Sentry and Grafana used about 3K each. Jira alone was about 17K. Tool count is a weak guide: long descriptions and big schemas cost more than many small tools, so check the token figure, not the number of tools.

Does too many tools hurt accuracy too?

Yes, by Anthropic's internal tests. With large tool libraries, Opus 4 scored 49% on its MCP evaluations with every tool loaded and 74% with tool search on; Opus 4.5 went from 79.5% to 88.1%. These are Anthropic's own measurements on its own models, so treat them as a direction rather than a rule for every client.

Do Claude Code, VS Code and Cursor limit MCP tools?

They handle it differently. Claude Code turns on tool search by default on supported models, loads only tool names and server instructions at start, and says it has no fixed per-server tool cap; the practical limit is your context window. It also warns when one tool's output passes 10,000 tokens and caps output at 25,000 by default. VS Code allows at most 128 tools per chat request. Cursor's current MCP docs state no tool cap; some older posts cite a 40-tool limit, so check your version. Using Claude Code? See the best MCP servers for Claude Code.

How do I cut MCP context use without losing tools?

Work through these in order, cheapest first. Each one keeps the tools you need within reach while removing the ones the current job never calls.

1. Disable servers per project. Keep a per-project config with only the servers that project uses, and switch the rest off in the client's tools picker.
2. Load toolsets, not whole servers. Servers like GitHub's let you enable named toolsets; its README says enabling only what you need helps tool choice and reduces context size.
3. Turn on tool search. Where the client supports it, defer tool definitions so only names load at start and full schemas load on demand.
4. Cap tool output. Limit how much a single tool result can return (Claude Code uses MAX_MCP_OUTPUT_TOKENS) so one big response doesn't flood the session.
5. Use code execution for multi-step jobs. Cloudflare's Code Mode and Anthropic's code-execution pattern let the agent call tools from code and read back only the final result; Anthropic's example fell from 150,000 to 2,000 tokens.

How do I choose which MCP servers to keep?

Keep the ones that are maintained, answer reliably and match a job you run every week. DartzDabz lists 1,576 MCP servers and scores 1,325 of them daily on trust and uptime (counts as of Oct 24, 2026, DartzDabz directory), so you can compare candidates before you add one. (browse all MCP servers) Prefer an official server for each system, drop duplicates that cover the same app, and retire any server the agent hasn't called in a month. How our own crew keeps each bot's tools short is in the DartzDabz playbook.

Built a lean MCP server? Submit it for a trust score.