Skip to content 
 _ __ ___  _   ___  _____  __| |
| '_ ` _ \| | | \ \/ / _ \/ _` |
| | | | | | |_| |>  <  __/ (_| |
|_| |_| |_|\__,_/_/\_\___|\__,_|

Claude Code

Claude Code is one of the most capable coding agents available, but its performance degrades as you add MCP servers. muxed fixes this by keeping tool schemas out of the context window and letting the agent discover tools on demand.

Setup

Section titled “Setup”
Terminal window
npx muxed init

init discovers your existing MCP servers and injects CLI instructions into ~/.claude/CLAUDE.md. See Setup for details and options.

Why muxed matters for Claude Code

Section titled “Why muxed matters for Claude Code”

Every MCP server you add to Claude Code dumps its full tool schema into the context window. This has compounding costs:

  • 20-30% of context consumed by tool schemas before the agent starts working. That is context you paid for that produces no useful output.
  • Skills, prompts, and default tools are deterministic — models always execute them. MCP tools compete for attention. The more you add, the less your designed trajectories get followed.
  • Tool selection accuracy drops as the number of available tools increases. The agent picks the wrong tool more often, leading to wasted calls and failed tasks.
  • Offloading tools to muxed is context engineering at the infrastructure level. Your carefully crafted instructions get followed instead of being drowned out by tool schemas.

What changes with muxed

Section titled “What changes with muxed”

Consider a concrete example: a user asks the agent to fix a customer query issue. The agent has access to a dozen MCP servers — postgres, filesystem, analytics, monitoring, and more.

Without muxed, all tool schemas sit in the context window from the start. The agent tries to figure out which tools to use from the crowded context. It picks the wrong ones — postgres/execute_query to run arbitrary SQL, filesystem/read_file to scan log files — and spends 38+ tool calls chasing dead ends. The “customer-analysis” skill that was specifically designed for this workflow never loads because the model’s attention is spread across hundreds of tool definitions. The task either fails or produces a fragile result.

With muxed, the context window is clean. The agent loads the “customer-analysis” skill first because there are no competing tool schemas pulling its attention elsewhere. When it needs a specific tool, it runs muxed grep "query", finds postgres/query_log, checks its schema with muxed info, and calls it directly. The issue is fixed in 3 steps.

The difference is not the model. It is the context. Fewer schemas means the agent follows the trajectories you designed instead of improvising with whatever tools happen to be visible.

Context window comparison

Section titled “Context window comparison”

Before muxed: Roughly 30% of the context window is consumed by MCP tool schemas. That leaves 70% for skills, prompts, and the actual task. Every tool definition competes with your instructions for the model’s attention, and adding more servers makes the problem worse.

After muxed: MCP schemas are gone from the context window entirely. That 30% is now free for skills, prompts, and conversation — the things that actually drive correct behavior. The agent follows your designed trajectories because there is nothing competing with them.

Next steps

Section titled “Next steps”